Personal tools
You are here: Home Events Past Events ANC workshop archive

ANC workshop archive

An archive of past ANC workshops — mainly internal talks

15/11/11  11:00 - 12:00

ANC Workshop Talks: Jyri Kivinen and Peter Orchard, Chair: David Reichert


Sparse structure in high-dimensional financial data

Finding structure in data is a fundamental problem in machine learning that has applications across multiple fields.

Standard approaches are often computationally expensive, and may not handle latent variables or high-dimensional data. In this talk, I will describe a method for learning a sparse, high-dimensional latent variable model, and illustrate its potential in financial stress-testing and visualising dependencies between asset prices.


Statistical Modelling of Natural Images: A Structured Approach

Natural images are effortlessly analyzed and parsed into semantically high-level descriptions by most humans. Producing such descriptions in image analysis systems is often complicated by the large and complex variability exhibited in the appearance of these high-dimensional signals, obtained via a noisy imaging process.

In this talk, I will discuss statistical modeling of such data using Boltzmann machines. Such Markov random field models with hidden units have shown significant promise for various unsupervised learning problems, including as models for statistical structure occurring in natural images.

We will begin by identifying critical issues with the current approaches, and argue that they are over-optimistic in terms of their goals. One evidence of this is the inability of the current generative models to produce image samples containing textured regions, a necessary subcomponent of any credible model for visual scenes. It is difficult to come up with quantitative assessment methods for unnormalized generative models; this can lead to rather distant proxies being used, such as discriminative performance of a classifier based on the generative model inferences.

Motivated by these observations we take a step back, and ask whether current methods are able to model even the more structured class of visual textures, and dissect them in the generative tasks of texture synthesis and inpainting, for which effective quantitative assessment methods are available. Based on the findings, we will also consider structured extensions to model more complicated visual data, starting by Boltzmann machines capable of generating multiple textures, and demonstrate state-of-the-art performance with them in texture modelling.

01/11/11  11:00 - 12:00

ANC Workshop Talks: Andy Gordon, Chair: Colin Mclean

Reverend Bayes meet Countess Lovelace: Probabilistic Programs, Semantics, and Tools for Machine Learning

Based on joint work with Johannes Borgstroem, U Uppsala; Michael Greenberg, U Penn; James Margetson, MSR; and Jurgen Van Gael, MSR.

The Bayesian approach to machine learning amounts to inferring posterior distributions of random variables from a probabilistic model of how the variables are related (that is, a prior distribution) and a set of observations of variables. There is a trend in Bayesian machine learning towards expressing generative models as probabilistic programs. As a foundation for this kind of programming, we propose a core functional calculus with primitives to draw from prior distributions and to observe variables. We define a rigorous semantics to our core calculus by introducing combinators for measure transformers, based on theorems in measure theory. Our system Fsoft implements our calculus by compiling a subset of the F# programming language to the Infer.NET machine learning library. Fsoft offers a higher level of abstraction than Csoft the native language of Infer.NET. The same Fsoft code may be executed to generate synthetic data and also compiled to factor graphs for statistical inference. We conclude by discussing future directions.

18/10/11  11:00 - 12:00

ANC Workshop Talks: Chris Gorgolewski and Cian O'Donnell, Chair: Matthias Hennig

Chris Gorgolewski

Using MRI to predict the outcome and optimize stroke rehabilitation

Stroke is one of the most common causes of disabilities, with 750,000 new cases each year in the US alone. In the following talk I will try to describe causes and symptoms of stroke and the recovery process. I will also try to outline research plan for using modern MRI techniques to evaluate and steer rehabilitation.


Cian O'Donnell

There's plenty of room at the bottom: stochastic ion channel gating in thin axons

Electrical activity in neurons is mediated by many small membrane proteins called ion channels. Although single ion channels are known to open and close stochastically, the macroscopic behaviour of populations of ion channels are usually approximated as deterministic. This is based on the assumption that the intrinsic noise introduced by stochastic ion channel gating is so weak as to be negligible. We have been examining cases where this assumption breaks down. We find that ion channel noise can mediate spontaneous action potential firing in small nerve fibres. We characterise the important factors and explore the possible implications for neuropathic pain disorders of peripheral nerves.

11/10/11  11:00 - 12:00

ANC Workshop Talks: Ali Eslami, Chair: Ian Simpson

Generative Probabilistic Models of Object Shape

When segmenting images of an object class, it can be very useful to have accurate information about the range of segmentations that are likely to be associated with that object class. These segmentations, or shapes, are binary images in which 'on' pixels indicate presence of the object of interest. Learning high-quality, probabilistic models of such images remains an open problem. The challenges are two-fold: 1) the images are typically large, 2) training data is hard to come by. In this talk, I will present a series of layered, undirected, probabilistic models in the Deep Boltzmann Machine (DBM) family which aim to overcome these two problems. I will also demonstrate applications of our models to medical and Kinect datasets.

Joint work with Nicolas Heess and John Winn (Microsoft Research).

27/09/11  11:00 - 12:00

ANC Workshop Talks: Mark van Rossum and Jim Bednar, Chair: Charles Sutton

Mark Van Rossum

Weight dependent synaptic learning rules

The strength of the synapses in the brain are presumably continuously subject to increases and decreases as the result of ongoing learning processes. This realization allows one to approximate the synaptic weight evolution as a stochastic process. This has been used to find fundamental limits of storage. Recently we introduced a synaptic information capacity measure based on Shannon information (Barrett and van Rossum '08). We use this to find the optimal weight dependent learning rules. We find that soft-bound learning rules are better than hard bound rules. Furthermore, we show how feedforward inhibition further increases storage.

Jim Bednar

How far can rate neurons go?

Cortical and thalamic neurons communicate via spikes rather than graded potentials, and a complete account of neural function will thus necessarily include spiking. However, the majority of experimental analyses (PSTHs, tuning curves, receptive fields, neural maps, etc.) and imaging techniques (optical imaging, two-photon imaging, etc.), are based explicitly or implicitly on the notion of a firing rate, which discards information about individual spikes. Accordingly, most of the aspects of visual neuron responses that have been clearly linked to visual function, such as preferences for orientation, motion direction, and spatial frequency, are expressed almost exclusively in terms of firing rates. Of course, many observations have been made of neural phenomena that depend crucially on spiking, but so far relatively few of these phenomena have been clearly relatable to visual function in real neurons, while providing information not present in the firing rate.

Currently, there are several huge, expensive multi-site EU, UK, or US projects to build massive spiking models of sensory cortex, consuming vast quantities of computational power and having vast quantities of underconstrained parameters. Alongside such simulations, it might be worthwhile to build firing-rate models that are similarly ambitious in scale but much simpler to specify, debug, analyze, build, and simulate. Such models will at worst serve to demonstrate clearly what added functionality is provided by the spikes, beyond the spiking rate, which might help focus the spiking-neuron modelers into clearly demonstrating the "added value" from spiking for cortical function. In this talk I'll present a specific model of temporal processing in the visual cortex under development with Jean-Luc Stevens, investingating how close a match to typical experimental results can be obtained without actually simulating the spikes, and discuss where we might want to take this model.

06/09/11  11:00 - 12:00

ANC Workshop Talks: Matthew Down and Nestor Milyaev, Chair: David Sterratt

Matthew Down

Combing multi-electrode array (MEA) recording of the developing mouse retina with light stimulation

In this talk I will cover what I've been doing during the first few months of my post-doc at Edinburgh/Newcastle universities. In particular I shall talk about setting up an MEA to receive light stimulation, the preliminary data and the types of experiment we can do with this set-up. I shall also talk about work I've done deploying services on the Carmen portal, a site for storing and analysing electro-physiological data.


Nestor Milyaev

Virtual Fly Brain, an online interface for 3D data browsing

Complex 3D datasets of biological information are most commonly handled as series of sections. These data are expensive to generate and thus their sharing and distribution is important. We present an online interface for browsing 3D stacks of biological images that allows easy handling of data several gigabytes in volume. To make such image data useful, mechanisms for linking it with underlying organism ontology and tools for querying anatomy and experimental databases are also provided.

Please note change of venue to IF G.07A


30/08/11  11:00 - 12:00

ANC Workshop Talks: Ronald Begg, Chair: Hugh Pastoll

CO-Releasing Molecules/Gene-Regulatory Network Inference

Carbon Monoxide (CO) has well known harmful effects.  However CO is produced within the body and has been shown to have beneficial effects such as suppressing organ-graft rejection in rats, mitigating tissue inflammation and inducing vasodilation.  Carbon Monoxide Releasing Molecules (CORMs) are of interest due to the potential of localised delivery of CO where it might be needed.

I will talk briefly about CO, CORMs, and the ongoing work being done in Sheffield on CORM-3, a water soluble CORM.

In addition I will discuss some modeling/inference ideas for gene-regulatory networks which have been considered in connection with the CORM-3 project.  The desire is to have a block-structured network where each block has a similar substructure to the others.  I will describe what has been attempted so far and possible ways forward.

23/08/11  11:00 - 12:00

ANC Workshop Talks: Botond Cseke, Chair: Frank Dondelinger

Bayesian Source Localization with the Multivariate Laplace Prior

I will present a 2009 NIPS paper with the same title by M. van Gerven, B. Cseke, R. Oostenveld, T. Heskes.

We introduce a novel multivariate Laplace (MVL) distribution as a sparsity promoting prior for Bayesian source localization that allows the specification of constraints between and within sources. We represent the MVL distribution as a scale mixture that induces a coupling between source variances instead of their means. Approximation of the posterior marginals using expectation propagation is shown to be very efficient due to properties of the scale mixture representation.  The computational bottleneck amounts to computing the diagonal elements of a sparse matrix's inverse. Our approach is illustrated using a mismatch negativity paradigm for which MEG data and a structural MRI have been acquired. We show that spatial coupling leads to sources which are active over larger cortical areas as compared with an uncoupled prior.

09/08/11  11:00 - 12:00

ANC Workshop Talks: Peggy Series and Amos Storkey, Chair: Mike Smith

Peggy Series

Learning and Unlearning of Perceptual Priors

In this talk, I will describe how perceptual expectations, often modeled as Bayesian priors, are thought to influence perception, as well as how they can be updated through statistical learning. In particular, I will review recent experimental work from my team, showing how expectations about motion direction can be quickly learned, leading to perceptual biases and hallucinations in human observers (Chalk, Seitz & Series 2010), as well as current extensions of this work to address the issue of transfer of such perceptual priors to similar objects (MSc project of Nikos Gekas). I will also present recent work showing that the prior belief that objects are static or move slowly rather than fast (Weiss, Adelson & Simoncelli, 2002), which is thought to reflect the statistics of natural stimuli and to explain a number of visual illusions, can be quickly unlearned and inverted (Sotiropoulos, Seitz & Series, in revision).


Amos Storkey

Machine Learning Markets

In this talk I will state what I think are 3 big issues in machine learning, and introduce the idea of machine learning markets, along with some of the developments that result from this basic idea. Machine learning markets involve using prediction market mechanisms for doing machine learning. One result of this is that implicit probabilistic models can be defined that extend some of the standard ways of combining machine learning components together. I will try to be fairly introductory, but as a result might annoy some people by not rigorously working through all the equations. Poor them.

02/08/11  11:00 - 12:00

ANC Workshop Talks: Athina Spiliopoulou and Xavier Oliver, Chair: Jyri Kivinen

Presentation by Xavier Oliver

Title - Assessing cell double labeling in the thalamus


In the developing brain, and specifically the thalamus, cells express varying combinations of genes over time that condition their location and fate. In this talk I will describe a current project that measures levels of cell double labeling in in situ data from the Allen Developmental Brain Atlas.

Presentation by Athina Spiliopoulou

Title - What are the open problems in Bayesian statistics?


For his column in the March 2011 bulletin of the International Society for Bayesian Analysis, Michael Jordan conducted a poll, where he asked a group of senior statisticians to list the top two or three problems in Bayesian statistics. In this talk I will go through the results of this poll, which are organised into a top-five list. I will describe the problems and try to set the ground for further discussion.


26/07/11  11:00 - 12:00

ANC Workshop Talks: Nigel Goddard and David Willshaw, Chair: Peter Orchard

Presentation by Nigel Goddard

Title: Topic Modelling with Blog Data

Abstract: I will present the outline of a research project to use probabilistic methods to answer questions about blogs and blog posts. Topic modelling using models using models based on Latent Dirichlet Allocation have become popular in analysis of textual corpora. I will introduce standard LDA and several variants, and discuss which may be most appropriate for answering particular kinds of questions about corpora of blog posts.

Presentation by David Willshaw

Title: The International Neuroinformatics Coordinating Facility.

Abstract: The INCF is an international organisation with a mission to coordinate and develop neuroinformatics worldwide. Sixteen countries take part. The MRC provides the UK governmental representation and I am the scientific representative. I am also the Coordinator of the UK Node of the INCF. In this talk I will give an introduction to the structure and activities of the INCF and will point out possibilities for involvement.


12/07/11  11:00 - 12:00

ANC Workshop Talks: Betty Tijms and Douglas Armstrong, Chair: Chris Gorgolewski

Betty Tijms

Morphological networks in the Edinburgh High risk of Schizophrenia Study

Previously I have presented a new method to extract individual morphological networks from grey matter magnetic resonance images (MRI). This method was developed and tested in a sample of healthy individuals, and network properties were studied for the first time with tools from graph theory. All the networks were small world and had property values comparable to those previously described in other morphological and functional MRI studies.

In this talk I will present the preliminary results after applying this method to a clinical sample: the Edinburgh High Risk of Schizophrenia study.


Douglas Armstrong

SynSYS and Brainwave - combining systems biology and invertebrate neuroscience to reduce reliance on mammalian in vivo experiments

SynSYS is an EU, FP7 funded consortium looking at the genetics of synaptopathies - disorders of the nervous system that can be linked to molecular events in the synapse. At Brainwave we translate the systems biology analysis of these synaptopathies into new humanised Drosophila reagents which can be used to help understand the molecular mechanisms of disease. I will present the background to the approach as well as recent results.

28/06/11  11:00 - 12:00

ANC Workshop Talk: Colin Mclean, Chair: Ondrej Mandula

Modularity and community structure in static networks

The application of unsupervised machine learning techniques to static real world network has led to the development of some important algorithms, that have help discover their underlying structure. The networks of interest include social, computer and biological networks, which have been found to exhibit community structure.

Protein-protein interaction data sets supplied by the large EU consortium's of EuroSPIN and SynSys, consisting of about 2100 proteins from the mammalian pre synaptic and post synaptic densities, are becoming available and need to be analysed. This talk aims to introduce two important community finding algorithms, which have been implemented and will be developed in preparation for this data.

21/06/11  09:15 - 17:00

ANC Review Day

14/06/11  11:00 - 12:00

ANC Workshop Talks: Christopher Ball and Matthias Hennig, Chair: David Acunzo

Christopher Ball - Modeling the development of color maps in macaque monkey

In macaque primary visual cortex (V1), color-selective cells are found in small, spatially separated "blobs", in contrast to orientation-selective cells, which form spatially contiguous maps. Recent optical and 2-photon calcium imaging studies of color-selective cells in V1 indicate that each color blob contains a range of hues, and that perceptually similar hues are adjacent, suggesting that the organisation reflects important features of subjective color perception rather than of physical wavelengths. In this talk, I will go through the experimental data, and present my results from trying to model the self-organization of color and orientation maps.

Using simulated LMS cone activations in response to natural scenes as training data, I find that color blobs develop into two distinct categories: one type for L/M cones, and another type for S cones. L and M cone responses are highly correlated because of their overlapping wavelength sensitivity functions, whereas S cone responses are only weakly correlated with L or M responses. If instead I simulate cone activations based on more uniformly separated sensitivity functions (e.g. camera RGB receptor sensitivities), and I equalize the hues present in the training images (to eliminate bias in the distribution of colors present in natural image databases), I find the model develops blobs selective to all hues. I'll discuss how these findings relate to the experimental results, and what they might tell us about how color processing circuitry develops.

Matthias Hennig - How to analyse a million spikes

I will talk about some recent work on analysing spiking data from high-density multielectrode array recordings. Joint work with Dagmara Panas, Oliver Muthmann and Luca Berdondini (IIT, Genova).

07/06/11  11:00 - 12:00

ANC Workshop Talks: Guido Sanguinetti and Hugh Pastoll, Chair: Lysimachos Zografos

Guido Sanguinetti - Signal processing for point processes

I will describe some applications of signal processing techniques to event based data in time or space-time. The basic framework is provided by log-Gaussian Cox processes, where we explicitly model the (discrete time) dynamics of the latent Gaussian field by using state-space models. I will describe two applications of this approach: processing neural spike data, and spatio-temporal modelling of the Wikileaks Afghan war diary data. All the work was done by Andrew Zammit-Mangion, a PhD student who was visiting late last year.

Hugh Pastoll - How to play advanced snake: in-vivo optogenetic control of unrestrained C. Elegans

Direct, global and rapid control of neuronal activity in normal behavioural conditions provides a unique opportunity to investigate the relationship between brain and behaviour. Here I discuss a new technique, incorporating optogenetic methods and automated worm tracking, which affords fine-grained real-time control of C. Elegans behaviour.

Reference: Andrew Leifer et al. 2011. Optogenetic manipulation of neural activity in freely moving Caenorhabditis elegans. Nature Methods

10/05/11  11:00 - 12:00

ANC Workshop Talks: Bilal Khan and Ondrej Mandula, Chair: Alex Mantzaris

Bilal Khan - Modelling Maps in the First Division of Auditory Cortex

Experiments in natal rewiring between the auditory and visual cortices of newborn ferrets suggest that mammalian auditory and visual systems are similar to the extent that rewiring at birth still produces functional maps in adult animals. Given these observations we hypothesise that these two sensory perception systems share a similar developmental process capable of regular adaption to either stimulus. Furthermore, recent experimental results suggest that the first division of auditory cortex contains spatial maps for sound frequency and that these maps depend on the auditory environment during development, just as found previously for the visual cortex.

I am employing the Topographica neural map simulator, originally designed to model the human primary visual cortex, to model those transformations performed by stations of the ascending auditory pathway: ultimately feeding a cortical layer by explicitly modeling the responses of the the primary and ultimate stations of subcortical processing (and implicitly incorporating those stations in between). 

Ondrej Mandula - Machine learning meets optical microscopy

Resolution of the optical microscope is limited by diffraction. In the last decade several fluorescence microscopy techniques have been proposed (making use of multiple images) to yield sub-diffractionresolution.

We are using machine learning methods such as Nonnegative Matrix Factorisation for data from a fluorescence microscope. The samples are labeled with flourophores exhibiting intemittancy (blinking) over time. By analysing the time series of the data we can recover individual fluorophores (sources) and localise them with a precision higher than the diffraction limit.

This can provide experimentally simple and non-expensive high resolution microscopy method.


03/05/11  11:00 - 12:00

ANC Workshop Talks: David Reichert and Richard Shillcock, Chair: Xavier Oliver Duocastella

Please note the change of venue to G.07

David Reichert - Deep Boltzmann Machines as Generative Models of Object-Based Attention in the Visual Cortex [practice talk]

I will be reviewing my work on modelling cortical perceptual phenomena with Deep Boltzmann Machines, with a focus on attentional processing: Here, hierarchical recurrent processing allows the model to retrieve information about objects in cluttered images, utilizing generatively learned knowledge about object shapes. This is a short practice talk for a conference (15th ICCNS), so feedback will be very welcome!

Richard Shillcock - New data on what's going on inside a dyslexic fixation

(with Mateo Obregon, Hamutal Kreiner &  Matthew Roberts)

We present initial analyses of the movements of the two eyes within individual fixations during the reading of text. The data are from a new ESRC-funded Edinburgh Corpus of Dyslexic Eye-movements in Reading, compiled at the end of 2010. Sixty-three students classed as dyslexic were tested on a range of language and cognitive tasks and were eye-tracked as they read 5000 words of text. The initial analysis of the binocular fixation data show consistent differences between typical readers (from an existing corpus of eye-movements acquired in the same way) and the dyslexics. Both sets of data contain interesting new phenomena relevant to an understanding of reading based on the anatomy of the visual pathways.

19/04/11  11:00 - 12:00

ANC Workshop Talks: Lysimachos Zografos and David Acunzo, Chair: Athina Spiliopoulou

Lysimachos Zografos - Agents provocateurs in the synaptic proteome of D. melanogaster

In this talk I will give an overview of the project that I'm currently involved in, working for Brainwave Discovery, as part of SynSys. We have designed and currently implementing a pipeline of high throughput behavioural screening of humanised fly lines. In this study, as a first stage, we will to introduce novel agents (human genes) in the fly synaptic proteome and use the Brainwave assay to test for effects in synaptic function as well as molecular biology and imaging and bioinformatics methods to elucidate mechanisms in potential candidates.

David Acunzo - Very early object and facial processing is modulated by spatial and object-based attention

I am going to present the data of 3 ERP experiments designed to test the effects of spatial and object-based attention on very early emotional facial expression processing (50 - 100 ms post-stimulus onset). The data suggests that both types of attention have an effect in this early time window. Interestingly, the effect does not seem to be face-specific, and could possibly be related to object identification.

12/04/11  11:00 - 12:00

ANC Workshop Talks: Mike Smith and Frank Dondelinger, Chair: Seymour Knowles Barley

Mike Smith - Finding evidence for Head and Place activity using fMRI

Place and Head Direction cells have been discovered in several non-human animals. It seems reasonable to assume they are also to be found in humans. fMRI studies using virtual reality navigation paradigms have reported that place is decodable from the activity in the hippocampus. My experiment last year attempted to reproduce these results and extend them to the decoding of head direction from the retrosplenial region. Although some aspects of the stimulus could be decoded, place and head direction seemed largely elusive. Many methods and options are available for decoding. This talk is aimed at eliciting suggestions on how to proceed with the research.

Frank Dondelinger - Inferring Gene Regulatory Networks using Linear Regression with Changepoints [DREAM 5 Redux]

Together with collaborators, I participated in the Network Inference Challenge of the DREAM 5 ( DREAM is project to test the predictive power of mathematical models in systems biology by getting teams to compete against each other in set challenges. The goal of this particular challenge was to infer four gene regulatory networks, with a total of almost ten thousand genes, from gene expression data. We applied two models, one based on L1-penalised linear regression, and the other based on Bayesian regression with changepoints inferred using RJ-MCMC. I will pick up where my last ANC talk left off and present the results of the DREAM 5 competition, explain how our team performed and why, and present some ideas for tackling this challenge in the future.

05/04/11  11:00 - 12:00

ANC Workshop Talks: Peter Orchard and David Sterratt, Chair: Matthew Down

Peter Orchard - Learning the Covariance of Financial Data

Modelling the distribution of stock returns is important for many financial applications. I will describe the portfolio selection problem, illustrating the need to model the stock return covariance matrix. I will give an overview of a LASSO-based approach to modelling the covariance of a Gaussian random field, and present preliminary results from this model.

David Sterratt - The development of ordered maps from the retina to its targets

During late prenatal and early postnatal neural development in vertebrates the axons from retinal ganglion cells grow and are pruned so as to form a topographic mapping from the retina to its target regions. Various experiments demonstrate that two broad classes of mechanism are necessary for a correct mapping to form: expression of marker molecules in the retina and the target regions, and mechanisms dependent on neural activity. However, exactly how these mechanisms operate and interact with each other is not known. There are a number of hypotheses, some of whose predictions have been derived rigorously using mathematical and computational models.

In this talk I will give an overview of our current understanding of this topic.

22/02/11  11:00 - 12:00

ANC Workshop Talks: Jyri Kivinen and Andrew Dai, Chair: Jakub Piatkowski

Jyri Kivinen - Transformation Equivariant Boltzmann Machines

In this talk I will describe a novel modeling framework for Boltzmann machines. The framework enables the inferences of the models to transform in response to transformed input data in a stable and predictable way, and avoids learning of multiple features differing only with respect to the set of transformations. Extending prior work on translation equivariant (convolutional) models, we develop translation and rotation equivariant restricted Boltzmann machines and deep belief nets, and demonstrate the effectiveness of these "steerable" models in learning frequently occurring statistical structure from artificial and natural images.

Andrew Dai - Learning about meetings

The automatic segmentation of meetings into phases where a group of related topics (such as agenda items) are being discussed is a challenging problem. I'll describe a model that models the lexical data available along with the topic profiles of each of the speakers in a meeting. This is done through a Markov model of the topic distribution of a meeting. I'll show preliminary results on newsgroup and ICSI meeting data.

08/02/11  11:00 - 12:00

Special ANC Workshop: Matteo Colombo and Matthew Chalk, Chair: Christopher Ball

Emergentist versus probabilistic approaches to understanding cognition

Which modelling approach is generally the most fruitful for explaining cognition? Two articles recently published in Trends in Cognitive Sciences debate the merits of approaching cognition from different ends of a continuum. On the side of *probabilistic modelling* we have Thom Griffiths, Nick Chater, Charles Kemp, Amy Perfors and Joshua Tenenbaum. Representing *emergentist* approaches are James McClelland, Matthew Botvinick, David Noelle, David Plaut, Timothy Rogers, Mark Seidenberg and Linda B. Smith. This contest is not short of "heavy-weights". Structured probabilistic takes a 'top-down' approach while Emergentism takes a 'bottom-up' approach. Both sides aim at understanding cognition. After having presented the (alleged) advantages and problems of both approaches, we will ask how substantial this debate is. In asking this question, we shall identify a number of issues for debate about how and when cognitive modelling helps us understand cognitive phenomena.

Griffiths, T., Chater, N., Kemp, C., Perfors, A., & Tenenbaum, J. (2010). Probabilistic models of cognition: exploring representations and inductive biases Trends in Cognitive Sciences, 14 (8), 357-364 DOI: 10.1016/j.tics.2010.05.004

McClelland, J., Botvinick, M., Noelle, D., Plaut, D., Rogers, T., Seidenberg, M., & Smith, L. (2010). Letting structure emerge: connectionist and dynamical systems approaches to cognition Trends in Cognitive Sciences, 14 (8), 348-356 DOI: 10.1016/j.tics.2010.06.002

25/01/11  11:00 - 12:00

ANC Workshop talks: Ali Eslami and Adrianna Teriakidis, Chair: Betty Tijms

Ali Eslami - Hierarchical Probabilistic Models for Object Segmentation

One of the long-standing goals in machine vision has been the task of foreground/background or object segmentation, in which an image is partitioned into two sets of pixels: those that belong to the object of interest in the foreground, and those that do not. I'll review some of the literature in this area, and present a generative probabilistic model that uses parts-based representations of object classes to perform this task.

Adrianna Teriakidis - Two papers about decision making

I will present the following two papers which attempt to influence decision making through brain stimulation:

Microstimulation of macaque area LIP affects decision-making in a motion discrimination task. Hanks TD, Ditterich J, Shadlen MN. Nat Neurosci, 2006 9:682-689.

Disruption of the right temporoparietal junction with transcranial magnetic stimulation reduces the role of beliefs in moral judgments.
Young L, Camprodon JA, Hauser M, Pascual-Leone A, Saxe R.
Proc Natl Acad Sci U S A. 2010 Apr 13;107(15):6753-8.

18/01/11  11:00 - 12:00

ANC Workshop talks: Chris Gorgolewski and Iain Harlow, Chair: David Reichert

Chris Gorgolewski - Computational challenges of using functional MRI and Diffusion Tensor Imaging for planning brain tumour extraction surgeries.

For malignant brain tumours are usually treated by partial or full resection. This procedure is aimed at maximising the amount of oncologically eloquent tissue being resected while preserving patients cognitive skills and quality of life. In this talk I will show how modern MRI techniques can help to achieve this goal, how they compare to currently used methods and most importantly what data processing and modelling challenges they present.

Iain Harlow - Neural models of recognition memory: Consensus, controversy and future directions

Episodic and, in particular, recognition memory has been heavily studied from the level of individual neurons to human behaviour, and in recent years attention has turned more and more towards producing a global theory of the neural and functional organisation of recognition memory in humans. Such an endeavour necessarily requires integrating and explaining evidence from many disparate sources, including (but not limited to) animal models, computational models, single cell recordings, lesion studies, imaging and human behaviour. I will outline some of the challenges this presents, and how researchers intend to address these. I will also focus on three questions in particular, the answers to which are fundamental to a global understanding of recognition memory, but which have been the source of considerable recent dispute: (1) How many processes support recognition, (2) What is the nature of these processes, and (3) How do these processes map onto brain structures?

My overall aim is to provide the non-expert a brief overview of the "state of play" in this interesting field, and specifically to highlight what we believe we know about recognition memory, as well as the questions which need to be addressed over the coming decade.

23/11/10  11:00 - 12:00

ANC Workshop talks: Jakub Piatkowski and Cian O'Donnell, Chair: Iain Harlow

Cian O'Donnell - Tuning synaptic integration for neural information processing

Neurons are often separated into different types based on their anatomical and physiological properties. However, individual neurons within a given cell type can also show substantial differences from each other. This within-type heterogeneity is usually attributed to random biological variability. However, recent data suggests that this variability might instead play a role in optimising neural information processing. This is an open problem. I'll review some of the experimental literature, compare the current theoretical models, and discuss implications for neural computation.

16/11/10  11:00 - 12:00

ANC Workshop talks: Alex Mantzaris and Matthew Down, Chair: Andrew Dai

Matthew Down - Modelling axonal pathfinding at the developing mouse optic chiasm.

Healthy pre-natal development of the visual system requires that the projection of retinal ganglion cell axons from the retina follow distinct trajectories to their targets in the thalamus and superior colliculus. The precise nature of these trajectories at the mouse optic chiasm and their mathematical relation to gradients of molecular guidance cues is the focus of my PhD. My talk will be an update on progress so far.

26/10/10  11:00 - 12:00

ANC Workshop talks: Betty Tijms and Seymour Knowles Barley, Chair: Nestor Milyaev

Betty Tijms - Quantifying the gray matter structure of individual brains in health and disease

Describing gray matter MRI data in a concise way to compare individual brains is not a trivial task. Previously I have presented a new method based on similarity of gray matter structure to extract networks from individual gray matter MRI data. The network description offers the advantage to describe the brain with statistics from graph theory.  This way, the network statistics of individuals or groups can be compared, while leaving the individual brains in their native space.  The method was applied to a sample of healthy volunteers, and was shown to extract networks with property values in line with previous research.

To investigate whether the method will show differences for a different sample of people from the test data, I have applied it to a cohort of aging people (n = 30, mean age = 76 ), in which some have developed Alzheimer?s disease (n= 10 ) and some show mild cognitive impairment (n=10). It is expected that the network property values would be different from the younger healthy sample (mean age = 35 ) as the morphology and size of the brain changes during aging. Also it is expected that within aging group the networks from people with impairments or Alzheimer?s disease will show different properties values than the networks from healthy people.

In this talk I will give a brief overview of the method, and a presentation of preliminary results.

Seymour Knowles Barley - Receptive Fields for EM Image Alignment and Neural Circuit Reconstruction

Serial section transmission electron microscopy (ssTEM) images of neural tissue can produce very high resolution reconstructions of neural morphology, including synaptic detail and in some cases protein localisation. Alignment and reconstruction of ssTEM images is currently performed manually or semi-automatically, with the aid of computer software, to generate a 3D model of the imaged neural circuitry. In some cases approximate alignment can be achieved automatically but high quality circuit reconstructions still require many hours of manual annotation.

Here 2D receptive fields, similar to those found in biological vision systems, are learnt from ssTEM data using supervised learning techniques. These receptive fields are applied to ssTEM images to automatically annotate neuronal membrane, synaptic connections, and organelles such as mitochondria. Objects recognised by the system can be used to improve alignment of serial images and produce partial 3D reconstructions as a starting point for further manual annotation.

05/10/10  11:00 - 12:00

ANC Workshop talks: Amos Storkey amd Richard Shillcock, Chair: Irina Erchova

Richard Shillcock - Work carried out collaboratively with Morten Christiansen, Louise Kelly and Sally Greenfield; paper in press in Cognition

I will report a study in which we tested stable Broca's aphasics for their capacity to respond to an Artificial Grammar Learning (AGL) task. In the AGL paradigm, subjects typically see short strings of letters presented in isolation on a screen, which they are instructed simply to observe.  They are later tested on different strings constructed from the same underlying finite-state grammar, versus similar but "ungrammatical" strings.  Typically, subjects can be shown to have "induced" the underlying grammar.  I will report whether this learning holds for Broca's aphasics, who are characterized by limited, dysfluent language production, and I will discuss the implications for future work with the AGL paradigm, including a second, recent unpublished study.

28/09/10  11:00 - 12:00

ANC Workshop talks: Frank Dondelinger and Hugh Pastoll, Chair: David Sterratt

Frank Dondelinger - Lessons from DREAM 5

Together with collaborators, I participated in the Network Inference Challenge of the DREAM 5 ( DREAM is project to test the predictive power of mathematical models in systems biology by getting teams to compete against each other in set challenges. The goal of this particular challenge was to infer four gene regulatory networks, with a total of almost ten thousand genes, from gene expression data. In this talk, I will review the issues that we faced, explain the benefits of our approach as well as its shortcomings, and share some of the lessons that we learned.

Hugh Pastoll: Dissecting neural circuits with light

Recent advances in optogenetics have established an exciting range of new tools for interrogating neural circuitry. I will discuss my endeavour to use ChannelRhodopsin2 to address hypotheses regarding the functional circuitry underlying grid cell formation in mouse entorhinal cortex.


07/09/10  11:00 - 12:00

ANC Workshop talks: Matthew Chalk and Athina Spiliopoulou, Chair: Matthias Hennig

Matthew Chalk - Does the brain speak 'Bayes'?

Over the past few years there has been increasing interest in probabilistic approaches to understanding vision, with the visual system interpreted as performing Bayesian inference in latent variable models. In support of this idea is a growing body of psychophysical evidence showing that people combine probabilistic information about their perceptual uncertainty with their prior knowledge in order to perform tasks optimally. However evidence that neurons encode information about probability distributions is, at this point, scarce. Indeed, while there are a great deal of models proposing different ways that neurons could encode probability distributions, a systematic comparison of how these models relate to each other, and what the key predictions are, is lacking. In this talk I will review some of these models, focussing in particular on the 'sampling hypothesis', where neural firing rates at each moment in time correspond to stochastic samples from the encoded probability distribution. Finally I will ask the question of how best (if possible) to turn the 'Bayesian brain' hypothesis into a clearly testable experimental prediction.

Athina Spiliopoulou - Probabilistic Models for Melodic Sequences

Learning a generative model for music is a particularly challenging task, as musical sequences exhibit not only local statistical relations, but also long-range and componential influences. In this talk, I will describe a Variable Length Markov Model and a Time-Convolutional Restricted Boltzmann Machine for modelling melodic sequences. I will show results on a dataset comprising melodies of reel tunes and discuss different evaluation metrics that can be used. Finally, I will point out possible extensions of the TC-RBM model and explain how these can increase performance on the musical task.

31/08/10  11:00 - 12:00

ANC Workshop talks: David Acunzo and Christopher Ball, Chair: Xavier Oliver Duocastella

David Acunzo - Is C1 modulated by facial expression?

After stimulus onset, the C1 component is the first visual evoked potential (VEP) recorded from scalp electrodes, generated by activity which is generally thought to be located in V1. Until recently, the C1 component did not attract much interest and was thought to be only sensitive to physical characteristics of the stimulus. More recently however, this view has been challenged and some evidence suggests that C1 can be endogenously modulated. In this presentation, I will talk about my latest experiment, testing the hypothesis of a modulation of C1 by the emotional content of the visual stimulus (facial expression). I will present the methods I used, and the results obtained.

Christopher Ball - Color in natural images and primary visual cortex

Perceptual color spaces represent the colors familiar to humans, arranged so that perceptually similar colors are close together (e.g. a circle of colors from red, through orange, yellow, and green, to blue, purple, and then back to red). Optical and two-photon imaging studies in macaque indicate that the organization of primary visual cortex (V1) might reflect these patterns of similarity, with cells organized into perceptual hue maps.

In order to see how these maps might arise, I am studying the distribution of hues in natural images and how the distribution changes at each level of the visual system. Starting from databases of calibrated color natural images, I'll show that the distribution of colors is extremely strongly biased, and is far from the uniform representation in perceptual color space. I am currently testing a variety of possible models for homeostatic adaptation in retinal photoreceptors and ganglion cells, to see their effects on this distribution, under the hypothesis that such adaptation mechanisms might balance out the unequal distribution over time.


24/08/10  11:00 - 12:00

ANC Workshop talks: Douglas Armstrong and Xavier Oliver Duocastella, Chair: Annie Huo

Douglas Armstrong - Genes after Cognition...

The Wellcome Trust Genes2Cognition (G2C) programme is a multi-centre, interdisciplinary research programme that started in 2002/2003. It is now in its wrap up phase but the work we have done in first building the first models of the post-synaptic proteome and then subsequently its expressional diversity, evolution and function is about to be continued through a series of followup studies. I will summarise what we aimed to do with G2C, what we achieved and what lies ahead for the new projects.

Xavier Oliver Duocastella - The Allen Developing Brain Atlas

In this talk I will describe the brain atlas of the developing mouse brain that was released earlier this year by the Allen Brain Institute. Some of the tools and methods to create it and that allow the study and comparison of gene expression patterns will also be explained.

17/08/10  11:00 - 12:00

ANC Workshop talks: Jim Bednar and David Reichert, Chair: Frank Dondelinger

Jim Bednar - A unified developmental model of maps, complex cells and surround modulation in the primary visual cortex

(Joint work with Jan Antolik)

Lateral connections between neurons in cat primary visual cortex layer 2/3 have been proposed to be involved in three very different phenomena: (1) surround modulation, i.e. interactions between separate visual stimuli, (2) topographic map development, e.g. for forming smooth orientation maps, and (3) formation of invariant responses, e.g. the phase invariance of layer 2/3 complex cells.  These phenomena have previously been studied separately, both in experiments and in models, and it is not clear whether the proposed roles for lateral connections are compatible and consistent.  Here we present a single, unified model that explains the development of orientation maps, formation of orientation and phase-tuned simple and complex cells in these maps, and surround modulation due to interactions between these cells.  In doing so, we not only offer a consistent explanation behind all these phenomena, but also create a very rich model of V1 in which the interactions between various V1 properties can be studied. The model allows us to formulate several novel predictions that relate the variation of single cell properties to their location in the orientation preference map in V1, and we show how these predictions can be tested experimentally.  Overall, this model represents a synthesis of a wide body of experimental evidence, forming a compact hypothesis for much of the development and behavior of neurons in the visual cortex.

David Reichert - Hallucinations in Charles Bonnet Syndrome Induced by Homeostasis: a Deep Boltzmann Machine Model

The Charles Bonnet Syndrome (CBS) is characterized by complex vivid visual hallucinations in people with, primarily, eye diseases and no other neurological pathology. We present a Deep Boltzmann Machine model of CBS, exploring two core hypotheses: First, that the visual cortex learns a generative or predictive model of sensory input, thus explaining its capability to generate internal imagery. And second, that homeostatic mechanisms stabilize neuronal activity levels, leading to hallucinations being formed when input is lacking. We reproduce a variety of qualitative findings in CBS. We also introduce a modification to the DBM that allows us to model a possible role of acetylcholine in CBS as mediating the balance of feed-forward and feed-back processing.  Our model might provide new insights into CBS and also demonstrates that generative frameworks are promising as hypothetical models of cortical learning and perception.


10/08/10  11:00 - 12:00

ANC Workshop talks: Charles Sutton and John Quinn, Chair: Hugh Pastoll

Charles Sutton - Supervision, But Not As We Know It

Typically, we divide learning into supervised and unsupervised approaches. Supervised learning systems can solve real practical problems, but they require manually labelled data, which can be expensive to obtain. Unsupervised learning algorithms need only unlabelled data, which is as plentiful as dirt, but it can hard to coax an unsupervised method into doing something that's actually useful. Recently, there has been an emerging theme of research that bridges this gap, by allowing people to pass information to learning algorithms in more fine-grained ways than simply labelling examples by hand. I'll discuss a few recent papers on this theme that I like; sadly, I didn't write any of them.

John Quinn - Spatiotemporal models for disease rate prediction with remote sensing data

In this talk I will discuss the spatiotemporal prediction of disease rates with probabilistic dynamical models. This modelling framework allows us to incorporate different sources of relevant observations such as remote
sensing satellite data and counts of related diseases, as well as less traditional sources of biosurveillance information such as pharmacy sales and school absenteeism rates.

03/08/10  11:00 - 12:00

ANC Workshop talks: Peggy Series and Matthias Hennig, Chair: Cian O'Donnell

Peggy Series - Drug addiction and abnormal decision making 

Drug addiction can be defined as the continued making of maladaptive choices.
In the last 10 years, a number of computational models have been proposed attempting to explain why an agent might continue to seek a drug or maladaptive behavior. These elegant models use the theoretical framework of reinforcement learning and are intimately linked with our understanding of the function of dopamine in the brain. They seem to have opened promising avenues for a growing field of research in modeling abnormal decision making. 
I will attempt to review this work.

Matthias Hennig - Age-dependent homeostatic plasticity of GABAergic signaling in developing retinal networks

Homeostatic compensation is a well-characterised phenomenon in neuronal cultures. Typically, such experiments demonstrate that, during sustained silencing or increases in electrical activity, individual neurons have the ability to adjust their intrinsic excitability and synaptic parameters to restore a certain firing rate. So far however, few studies have investigated homeostatic effects in intact brain circuits, and the effects on network activity are largely unknown. Here, I will present data from a series of homeostasis experiments in the intact mouse postnatal retina in vitro.  Spontaneous retinal waves were monitored with multi electrode arrays during prolonged GABAA receptor blockade with bicuculline. We found a compensation of spatio-temporal neural activity patterns at postnatal day (P) 5-6, when GABAA is depolarizing in retinal ganglion cells. Once GABAA responses switch to mature inhibition around P7, this ability for homeostatic adaptation disappears. These experiments suggest that homeostatic compensation depends on the state and activity of the whole neural circuit. Joint work with John Grady and Evelyne Sernagor at Newcastle University.


27/07/10  11:00 - 12:00

ANC Workshop talk: Iain Murray, Chair: Alex Mantzaris

Sampling latent Gaussian models and hierarchical modelling

What do modelling basketball scores and inferring dynamics of gravitational bodies have in common? I will outline some of my recent work[1] on practical Markov chain Monte Carlo inference in hierarchical models incorporating Gaussian processes. There will be some general messages about probabilistic modelling and constructing Markov chain Monte Carlo algorithms.


20/07/10  11:00 - 12:00

ANC Workshop talks: Richard Shillcock and Lysimachos Zografos, Chair: Peter Orchard

PLEASE NOTE CHANGE OF LOCATION FOR THIS WORKSHOP: Crystal McMillan Building, seminar room 1

Lysimachos Zografos - "Reconstructing models of the fly Postsynaptic Proteome "

In our effort to elucidate the architecture and mechanisms of molecular machines in the postsynaptic density (PSD) and how they're involved in cognition or disease, we apply a protein-protein interaction network (PPIN) reconstruction workflow on proteomics data. Although this approach has successfully been applied in mice, rats and humans (with an overlap of >80%) there was a serious lack of data for reconstructing a different (simpler?) model based on the PSD of another model organism in order to be able to compare this with the aforementioned models. In this talk I'll cover the first optimisation and application of such a proteomics and PPIN reconstruction workflow on D. melanogaster brain extracts and, present the reconstructed model of the fruitfly's PSD and discuss various issues concerning what was found and what was not found.


13/07/10  11:00 - 12:00

ANC Workshop talks: Vincent Danos and Mark van Rossum , Chair: Seymour Knowles Barley

Vincent Danos - Collective variables in biomolecular networks

We introduce a stochastic graph rewriting language to describe the dynamics of a soup of biomolecules - each possible interaction is captured by a local rewrite rule - the dynamics is obtained by giving rules rate constants and letting them compete. This gives rise to a continuous-time Markov chain.

Given the high degree of combinatorial-ness of the signalling systems we intend to describe, we ask whether it really makes sense to look at their dynamics in terms of the time courses of complex species. 

We propose another approach. By inspecting the rules governing the dynamics, we propose to compute “collective variables” that 1) are typically far fewer in number than complex species, 2) and that define a closed/self-contained differential system that exactly captures the global differential dynamics of the rule set (sometimes but not always also its stochastic dynamics.)

We explain the construction of these collective variables. Specifically, we explain how the average evolution of any local graph motif X can be described by a differential equation - where each term is related to the concentration of a certain type of glueing of X. We describe a saturation algorithm that constructs - given a seed, ie a motif of interest - a smallest set of motives with a self-consistent dynamics. (A scaleable version of this saturation algorithm is implemented - PNAS 106, 6453-6458, 2009).

We conclude by connecting this construction with some general discourse on emergence and learning - partly inspired by “The middle way” paper - PNAS 97, 32-37, 2000.

This is joint work with: Jerome Feret, Jean Krivine, Walter Fontana and Russ Harmer.


Mark van Rossum - Weight dependent synaptic learning rules

The strength of the synapses in the brain are presumably continuously subject
to increases and decreases as the result of ongoing learning processes. This
realization allows one to approximate the synaptic weight evolution as a
stochastic process. This has been used to find fundamental limits of storage
(Fusi and Abbott 2007)

Recently we introduced a synaptic information capacity measure based on
Shannon information (Barrett and van Rossum '08). We use this to find the
optimal weight dependent learning rules.
We find that soft-bound learning rules are better than hard bound
rules, although the improvement is quite small. Furthermore, we show how
feedforward inhibition further increases storage.

06/07/10  11:00 - 12:00

ANC Workshop talks: Robert Kyle and Graeme Phillipson, Chair: Nicolas Heess

Graeme Phillipson - An ant based algorithm for tracing neurons

As part of the DIADEM challenge project we have been trying several ideas for tracing neurons. The problem is to take microscopy images of neurons and trace the centre lines of all neurites. Most of the current algorithms work well on a limited set of image types, and do not generalise well. Here we will present one of the approaches that the team tried. The algorithm is loosely based on ant colony optimisation. With this approach many random walks are taken through a graph which represents the probability of different points in the image being connected. Ends of neurites are detected when a walk can go no further. Then the average path home to the seed point is traced out. The idea is that this replaces the difficult problem of detecting branches with the easier problem of detecting end points and then merging. Also this would allow for a confidence interval on the tracing to be calculated, which would be a big improvement over manual tracing.

Robert Kyle - What has computational modelling ever done for neuroscience?

During the course of my PhD I have constructed quite a few models, none of which worked as I had hoped.  What have I been doing wrong?The aim of this talk will be to describe those models I have made as part of my PhD, and look at the problems I encountered.  I want to talk about a theme that has been running through my thesis - what makes a good computational model? - and come up with some answers.  I'd like to hear yours too.As part of my cunning plan to see how many people read to the bottom of ANC seminar abstracts I would like you to come to the seminar with an answer to the question posed in the title.  Or more specifically, come with examples of what you consider to be successful computational models....


29/06/10  11:00 - 12:00

ANC Workshop talk: Guido Sanguinetti, Chair: Jakub Piatkowski

Guido Sanguinetti: Inference in continuous time stochastic processes

I will review some work on (mostly) variational approaches to perform inference in (mostly Markovian) continuous time stochastic processes which are (partially) observed at discrete time points. My main motivation lies in problems in Systems Biology, where mathematical concepts such as Markov Jump Processes play a prominent role, but there are several other possible application areas. I will first describe in some detail the way we perform variational inference for Markov Jump processes, and then move on to kinetic models (ODEs, SDEs) which are driven by a latent jump process.

22/06/10  11:00 - 12:00

ANC Workshop talks: Nicolas Heess and Iain Harlow, Chair: David Acunzo

Iain Harlow: Recollection has a threshold: Evidence from a novel source task.

Recollection, the retrieval of information and events from our past, can be rich and vivid, yet at other times desert us and leave us floundering for a friend’s name. A memory threshold has long been proposed to account for this: in some cases the threshold will be exceeded and the information retrieved, otherwise recollection fails completely. Alternatively, no threshold exists and recollection strength is a continuous, graded signal. A number of important neuroscientific results depend on the existence of this threshold – yet surprisingly, it remains hotly debated. Here we use a novel graded source task to clearly demonstrate that recollection is subject to a threshold. The results illuminate a fundamental (but previously unclear) characteristic of recollection, with important implications both for understanding how memory functions, and for measuring it accurately in the laboratory.

Nicolas Neess: Deep Segmentation Networks: Masked RBMs for factoring appearance and shape - and beyond

One hallmark of natural images is the variability of visual characteristics across different image regions and the presence of sharp boundaries between regions arising, for instance, from objects occluding each other. Even though Restricted Boltzmann Machines (RBMs) are capable of learning the structure of highly complex data without supervision they have difficultis modeling such sharp transitions from one set of image statistics to another. In my previous ANC workshop I described the Masked RBM, a model for image patches, which allows region boundaries to be modelled by factoring shape and appearance. The model employed an efficient representation of shape based on an explicit notion of (relative) depth. In this talk I will focus on bigger pictures: After a brief recap of he model for image patches I will describe how it can be extended to full images by considering a field of Masked RBMs. The field of Masked RBMs models images in terms of a large number of small, overlapping "objects", each of which has an associated shape and appearance. I will describe how inference and learning can still be done efficiently in this model and provide some results for toy data and natural images. Finally I will outline how the Field of masked RBMs naturally gives rise to a recursive, hierarchical framework for modeling images at different scales and levels of abstraction, the "Deep Segmentation Network", and I will discuss some of the challenges ahead.

Joint work with Nicolas Le Roux and John Winn (Microsoft Research).


08/06/10  11:00 - 12:00

ANC Workshop talks: Nigel Goddard and Chris Williams, Chair: Lysimachos Zografos

Chris Williams - Greedy Learning of Binary Latent Trees

Inferring latent structures from observations helps to model and  possibly also understand underlying data generating processes.  A  rich class of latent structures are the latent trees,  i.e.~tree-structured istributions involving latent variables where  the visible variables are leaves.  These are also called  hierarchical latent class (HLC) models. Zhang (2004) proposed a  search algorithm for learning such models in the spirit of Bayesian  network structure learning.  While such an approach can find good  solutions it can be computationally expensive.  As an alternative we  investigate two greedy procedures: the BIN-G algorithm determines  both the structure of the tree and the cardinality of the latent  variables in a bottom-up fashion. The BIN-A algorithm first  determines the tree structure using agglomerative hierarchical  clustering, and then determines the cardinality of the latent  variables as for BIN-G.  We show that even with restricting  ourselves to binary trees we obtain HLC models of comparable quality  to Zhang's solutions (in terms of cross-validated log-likelihood),  while being generally faster to compute.  This claim is validated by  a comprehensive comparison on several datasets.  Furthermore, we  demonstrate that our methods are able to estimate interpretable  latent structures on real-world data with a large number of  variables.  By applying our method to a restricted version of the 20  newsgroups data these models turn out to be related to topic models,  and on data from the PASCAL Visual Object Classes (VOC) 2007  challenge we show how such tree-structured models help us understand  how objects co-occur in images.

Joint work with Stefan Harmeling

Nigel Goddard  - Web-Based Computational Modelling

The increasing support provided for web-delivery of software provides opportunities for computational modelers.  The vision is of a modelling environment running on servers with a wide range of authoring, annotation, simulation, visualisation and analysis tools provided in the web-browser.  I will describe some components of such an environment under development for System Dynamics models. 


01/06/10  11:00 - 12:00

ANC Workshop talks: Nicolas Heess and Wan-Yu Hung, Chair: Robert Kyle

Cancelled due to DTC trip.

25/05/10  11:00 - 12:00

ANC Workshop talks: Cian O'Donnell and Roger Zhao, Chair: David Reichert

Roger Zhao - Modelling curve adaptation effects on high-level facial-expression judgments

Xu et al. (2008 J. Neurosci. 28:3374-83) recently showed that adaptation to curved lines can affect high-level perception of emotional affect in faces, apparently by modifying the perception of mouth curvature. For the patterns tested so far, the effect increases as the adaptation pattern curvature increases. However, standard low-level aftereffects like tilt and motion aftereffects typically show an S-shaped curve, with a decline in aftereffect strength for test patterns sufficiently different from the adaptation stimulus. In computational models of the TAE, the S shape reflects adaptation in neurons that prefer a particular orientation, with neurons preferring very different orientations exhibiting little adaptation and thus showing weaker aftereffects. To see whether a similar explanation might apply to curvature/emotion aftereffects, we adapted an existing model of the TAE for use with curved lines and added processing for emotion judgments. We found that the model replicated the existing data on high-level effects of low-level adaptation, but strongly predicted an S shape, i.e., that sufficiently large curvature values would lead to a lower effect. This prediction can be tested in humans, potentially helping to constrain the properties of the neurons underlying the effect.

Cian O'Donnell -Dendritic spines can stabilise synaptic strength

Although stable synaptic weights are important for long-term memory, it is not known how synapses retain their relative strengths over time. We've been working on a new model which might solve this problem.

Most excitatory synapses in the brain are hosted on small structures called dendritic spines. Spine size is tightly correlated with synaptic strength. The function of this relationship is not known. Notably, dendritic spine size is one of several factors that regulates calcium signalling at the synapse. Because calcium signals trigger synaptic plasticity, this spine size scaling might result in a weight dependence in the synaptic plasticity rule. We explored the implications of different spine-size to calcium-influx realationships on a synaptic plasticity rule and find that different scalings result in either stable or unstable synaptic weight dynamics. We then built a biophysical model of a spine and synapse using data from the literature and find that real synapses likely fall in to the 'stable' category. Finally, we explored each scenario's synaptic weight distributions and calculate synaptic lifetimes.

04/05/10  11:00 - 12:00

ANC Workshop talk: Nestor Milyaev, Chair: Matthew Down

Creating web interfaces for 3D imaging in biological applications

Nowadays, building web interfaces for 3D image viewing is much an ad-hoc business. In the scope of the Virtual Fly Brain project we aim at building a generic interface for browsing 3D images, such as a scanned tissue stack. The talk will outline the current state of play in the area, discuss possible solutions and alternatives approaches for solving the task. Then, the new viewer being developed for the VFB project will be presented; its features, possible difficulties and future plans will be discussed.

20/04/10  11:00 - 12:00

ANC Workshop talks: Jakub Piatkowski and Alex Mantzaris, Chair: Athina Spiliopoulou

Jakub Piatkowski - Discovering white matter structure in diffusion MRI images (an update on my PhD project).

I am developing methods for reliable estimation of white matter (WM) tract volume or coherence from diffusion MRI (dMRI) data. These parameters are necessary to study the condition of white matter in health and disease. Considerable amount of work in the recent years has focused on correct WM fibre angle estimation and tracking, which can now be done reasonably well. Tract integrity, however, is still estimated using an index (FA) derived from the Diffusion Tensor model, although it is well known that this estimate is not reliable. Worse still is the estimation of tract volume, for which there are no well established methods of acceptable precision. This project fills the gap by describing WM tracts completely, their physical extent as well as coherence. Information is shared across neighbouring voxels in order to exploit all dependencies present in the data.

Alex Mantzaris - "A transdimensional phylogenetic factorial hidden Markov model based on a codon-site specific nucleotide substitution model "

"The work is on improving the model of the Phylogenetic FHMM for detecting topology changes (recombination events), relative rates of evolution between codon positions, and rate heterogeneity along DNA sequence alignments.  The previous stage of this work required a specified number of rate states to be given and then were sampled along the DNA sequence alignment, which had the pitfalls that too many rate states would result in redundancy and too few would not capture the important features of the data. The number of rates states are now sampled via RJMCMC, which has been used previously, but without taking into consideration all the features of DNA modelled here in particular the codon structure. The results are presented in synthetic studies and a real data analysis." 

13/04/10  11:00 - 12:00

ANC Workshop talks: Adrianna Teriakidis and Annie Huo, Chair: Betty Tijms

Adrianna Teriakidis

I will give an overview of the three results sections of my thesis, which are all related to the development of the mammalian neuromuscular junction. These results could have consequences for understanding motor neuron development, central nervous system plasticity, degeneration and disease including motor neuron disease, other neurodegenerative disorders, injury, regeneration and the brain.

Annie Huo - "Adaptation and Neural Network for Sensory Integration"

Adaptation is the most important feature for Neural Networks. Here I would like to generally discuss adaptation in biology and AI at first.  An adaptive neural network for visual and auditory integration is introduced later, which is modelled for barn owl superior colliculus in my thesis. The following robot experiment and VLSI chip test shows this bio-inspired neural adaptation can work well in correcting the physical mismatch in sensor fusion. There are also many other popular methods for sensor fusion like kalman filter and bayesion network, a very brief description will be given. 

06/04/10  11:00 - 12:00

ANC Workshop talks: David Willshaw and Irina Erchova, Chair: Ondrej Mandula

David Willshaw - "Only Connect - again"

I am giving a shortened version of my Hamming seminar that I gave recently to the School - my apologies to those of you who heard the full version.

Amongst the topics I will discuss are:

(i) My research problem - modelling the development of nerve connections - and why it is important.

(ii) How my work fits into the work of the School of Informatics - in fact, what is Informatics?

(iii) My latest results and what they say about how the Eye connects with the Brain.

(iv) Some comments on some of Hamming's wise words that occur in the transcript of a talk he gave



Irina Erchova - Measuring Variability in Membrane potential oscillations

Many local networks are characterised by their oscillatory rhythm generated under various physiological and behavioral conditions. It can be crucial for defining movement precision, stimulus discrimination, and in attention and memory tasks. Small alterations in oscillatory behavior can cause severe pathology such as epilepsy. The network oscillations are driven either by mechanisms localized within individual neurons or by feedback interactions among populations of neurons. In individual neurons, oscillations often appear as membrane potential rise and fall that both limit temporal precision of action potentials and facilitating synchronous activity of neighboring neurons. However, generation of membrane potential oscillations (MPOs) and their role in shaping spiking output of a given neuron are not yet fully understood. The MPOs are generated via activation of voltage-gated channels and depend on their kitnetics and numbers, but co-activation of several channels often results in noisy fluctuations that difficult to predict. I am going to present some experimental data from the oscillatory cells in the medial temporal lobe and discuss variability in the oscillatory behavior in cells population, bias in measurements introduced by different recording techniques, and pharmacological experiments clarifying some of cellular mechanisms involved.

30/03/10  11:00 - 12:00

No ANC Workshop or Seminar

23/03/10  11:00 - 12:00

ANC Workshop talks: Matthew Down and Seymour Knowles Barley, Chair: Adrianna Teriakidis

Matthew Down - "Modelling axonal pathfinding at the developing mouse optic chiasm."

Healthy pre-natal development of the visual system requires that the projection of retinal ganglion cell axons from the retina follow distinct trajectories to their targets in the thalamus and superior colliculus. The precise nature of these trajectories at the mouse optic chiasm and their mathematical relation to gradients of molecular guidance cues is the focus of my PhD. My talk will be an update on progress so far.

16/03/10  11:00 - 12:00

No ANC workshop or Seminar

09/03/10  11:00 - 12:00

ANC Workshop talks: Athina Spiliopoulou and Betty Tijms, Chair: Wan-Yu Hung

Athina Spiliopoulou - Probabilistic Models for Generating Melodic Sequences

Learning a generative model for music is a particularly challenging task, since musical sequences typically exhibit not only local statistical relations, but also long-range and componential influences. In this work, we are interested in developing machine learning methods that are able to capture such complex relations directly from musical sequences, without the need to incorporate domain-specific knowledge. In this talk, I will motivate a Variable Length Markov Model and a Time-Convolutional Restricted Boltzmann Machine for tackling this problem and show results of the two models on a dataset comprising melodies of reel tunes. Finally, I will discuss possible ways of combining/extending these models to improve their performance on melody generation.

Betty Tijms - A new method to extract networks from individual grey matter MRI scans

Describing gray matter MRI data in a concise way to compare  individual brains is not a trivial task. Previous studies have  shown that the structure of the human brain can be represented as  an anatomical network, offering the advantage to describe the brain  with statistics from graph theory. In this talk I will present a  new method based on similarity of gray matter structure to extract  networks from individual gray matter MRI data. I will show that  this method offers a powerful and stable way to describe the  statistics of gray matter structure. Finally I will show  preliminary results that supports the hypothesis that similarity in  cortical structure between two brain regions indicates anatomical  connectivity between these regions. 

23/02/10  11:00 - 12:00

ANC Workshop talks: Hugh Pastoll and David Reichert, Chair: Xavier Oliver Duocastella

David Reichert - Deep Boltzmann Machines as Models of Top-Down Modulated And Attentional Object Perception in the Visual Cortex

I will introduce Deep Boltzmann Machines as models that could potentially shed light on some principles underlying cortical processing in general and visual processing specifically. In particular, I aim to show how these relatively simple models could make concrete a variety of different models and hypotheses about cortical processing such as Predictive Coding (Olshausen & Field '97, Rao & Ballard '99, etc.) and Hierarchical Bayesian Inference (Lee & Mumford '03). The suggested framework is furthermore based on the idea that the cortex implements perception one object at a time (Rensink '00), which relates the approach to models of attention (e.g. Tsotsos '08).

Hugh Pastoll - A spike clustering gradient along the dorsal-ventral axis of the medial entorhinal cortex reflects grid-cell spacing and predicts new conductances.

The density of the hyperpolarization-activated current (Ih) in principle stellate cells varies along the dorsal-ventral axis of mouse medial entorhinal cortex. We show that the Ih gradient supports a gradient in the active properties of stellate cells. A computational model of clustered spiking shows that features of clustered spike trains also suggest a gradient in the KCa conductance and predict a novel conductance. 

09/02/10  11:00 - 12:00

ANC Workshop talks: Christopher Ball and Frank Dondelinger, Chair: Alex Mantzaris

Christopher Ball - Development of Color Vision in Macaque Monkey

Human and macaque retina normally contains photoreceptors sensitive to three different bands of wavelength, termed the long (L), medium (M), and short (S) cones. These three cone classes sample the spectrum of light from a scene, and from this sampling the brain can create the perception of millions of colors.

In this talk, I will present recent experimental results about the development of macaque monkey's early colour vision system. I will focus on the wiring between retinal ganglion cells and photoreceptors (which appears to be unselective for L and M cones, in contrast to previous theory), and on how colour is represented in primary visual cortex (colour-selective cells are found in small, spatially separated "blobs", in contrast to orientation-selective cells, which form spatially contiguous maps). I will discuss recent optical and 2-photon imaging studies of colour-selective cells in V1 (indicating that each colour blob contains a range of hues, with perceptually similar hues being adjacent), as well as results of gene therapy to "cure" red/green colourblind squirrel monkeys.

Frank Dondelinger - Inferring Gene Regulation Networks using Heterogeneous Dynamic Bayesian Networks with Information Sharing

The classical paradigm for Dynamic Bayesian Networks (DBNs) is based on the homogeneous Markov assumption, where the edges between adjacent time points are time-invariant. Recently, several approaches for heterogeneous DBNs based on changepoint models have been proposed. Our approach extends an existing Reversible-Jump MCMC-based method by introducing a prior on network structures that allows for information sharing between network segments that are separated by a changepoint. I will present different forms that this prior can take, and show results on simulated data, as well as an application of the method to the problem of inferring gene regulation networks from gene expression time series data.

26/01/10  11:00 - 12:00

ANC Workshop talks: David Acunzo and Matthew Chalk, Chair: Iain Harlow

David Acunzo - Attentional orienting toward emotional stimuli

In human participants, visual attention can be modulated by the presence of an emotional stimulus in the visual field. I will present some behavioral results of work in progress, aimed at investigating exogenous and reflexive orienting of covert attention toward fearful faces. 


Matthew Chalk - Effects of attention on gamma frequency oscillations and spike field coherence in V1

I will be presenting work conducted in within the lab of Prof. Alex Thiele, looking at the effects of visual attention on the LFP signal in V1.

Attention increases gamma frequency synchronisation of V4 neurones representing the attended stimulus (Fries et al., 2001). Thus localised synchronisation changes might serve to amplify behaviourally relevant signals. Here we investigate whether these findings transfer to V1 of the macaque monkey. We measured the local field potential (LFP) and V1 spiking activity while monkeys performed an attention-demanding task. We show that gamma oscillations were strongly modulated by the stimulus and by attention. Stimuli that engaged inhibitory mechanisms induced the largest gamma LFP oscillations and the largest spike field coherence. Directing attention towards a visual stimulus at the receptive field of the recorded neurons decreased LFP gamma power and spike field coherence. This decrease could reflect an attention mediated reduction of surround inhibition (Sundberg et al., 2009). Thus, increased neuronal synchrony with attention is not a universal mechanism by which behaviorally relevant signals are amplified. In V1 neuronal synchrony depends on non-classical receptive field activation and attention mediated changes thereof.


12/01/10  11:00 - 12:00

ANC Workshop talks: Nigel Goddard and Xavier Oliver Duocastella, Chair: Andrew Dai

Nigel Goddard - Models of Residential Energy Demand

Residential needs account for 28% of UK energy demand.  In the context of UK comittments to reduce GHG emissions by 34% by 2020 and 80% by 2050, effective measures to reduce residential demand are required.  I will introduce some models of residential energy demand, focussing on agent-based models. 


Xavier Oliver Duocastella - Early development of the thalamus and the thalamocortical tract

The fibers connecting cerebral cortex and thalamus form a very important tract in the brain. Attempts at studying these connections at early stages of development will be described, and the current focus on the thalamic part of the system will be presented.

15/12/09  11:00 - 12:00

ANC Workshop talks: David Sterratt and Iain Harlow, Chair: Lysimachos Zografos

David Sterratt - Inference of original retinal coordinates from flattened retinae

In retrograde tracing experiments to determine the mapping of connections from the retina to the superior colliculus in mammals, fluorescent tracer is injected at a point in the superior colliculus and allowed to transport retrogradely down the axons of retinal ganglion cells to their cell bodies in the retina. The retina is then dissected and flattened, and the pattern of dye in cell bodies can be seen in the flattened retina. In the process of flattening the retina, a number incisions are made and tearing occurs. The pattern of label can be spread across incisions and tears in the flattened retina, complicating analysis of the mapping.

One way of simplifying the analysis would be to infer the positions of the cell bodies in the spherical coordinate system of the intact retina from their positions in the flattened retina. We present a method to achieve this approximately by minimisation of an energy function. A triangular grid is laid over the flattened retina. The coordinates of the grid are then transformed so that they lie on a partial sphere with the correct dimensions for the intact retina, and the transformation is then adjusted so as to minimise the sum of the squared differences between the lengths between corresponding pairs of adjacent points on the flattened and intact retinae. This also allows for lines of latitude and longitude in the intact retina to be projected onto the flattened retina.

This work has been carried out in collaboration with Daniel Nedergaard and Ian Thompson in King's College, London. I'm also grateful to members of ANC for their helpful comments when I spoke on the same topic this
time last year!

Iain Harlow - Recollection has a threshold: Evidence from a novel source task.

The majority of (but not all) memory theorists hold broadly to a multi-process view of recognition - that is, recognition can be supported by different processes, commonly familiarity, and recollection of context. There is general agreement that familiarity is graded, and well described by a gaussian signal-detection model. There is in contrast much debate over whether recollection is similar, or whether it contributes in an all-or-none manner (i.e. one recollects strongly, or not at all).

Most evidence (for both views) is based on receiver-operator characteristic (ROC) analysis, using confidence ratings to infer the underlying memory strength distributions. Here we take a different approach and examine the distributions more directly. We show that recollection has a clear threshold, making it incompatible with unidimensional signal detection models and providing further evidence that at least two functionally distinct processes contribute to episodic memory.

08/12/09  11:00 - 12:00

ANC workshop talks: Matthias Hennig, Chair: Graeme Phillipson

Matthias Hennig - Statistical models of multi-neuron spike data

Simultaneous recordings of long episodes of activity from many neurons with multi-electrode arrays provide a potentially rich source of information about neural circuit structure and function. Due to their size and complexity however, it is often challenging to process and analyse such data sets.  Recently, several studies have introduced parametric statistical models as a way of describing multi-neuron activity patterns. Here, I will review pairwise interaction maximum entropy models as the most parsimonious description of populations of coupled neurons, and discuss their advantages and limitations.

01/12/09  11:00 - 12:00

ANC Workshop talks: Cian O'Donnell and Andrew Dai, Chair: Edwin Bonilla

Cian O'Donnell - Dendritic spines can stabilise synaptic weights

Long-term memory requires stable synaptic weights. Most proposed solutions to this problem require elaborate and underconstrained molecular signalling mechanisms. We've been working on a simpler alternative model in which changes in dendritic spine size following plasticity can stabilise synaptic strength by modifying local calcium dynamics. We used a biophysical computer model of a spine to show how different calcium-influx to spine-size relationships can lead to stable, unstable or even bistable synaptic weight dynamics. When we use parameter estimates from the experimental literature, the model predicts that real synapses fall into the 'stable' category. We then built a reduced synaptic learning rule following either stable or unstable dynamics and explored the implications for synaptic weight distributions and memory storage in an integrate-and-fire neuron. This suggests a novel function for dendritic spines.

Andrew Dai

Entity resolution or record linkage is a difficult problem that is often the first step in data mining to reduce the noise in a dataset. When applied to authors and citation databases, this becomes a problem of discovering real-world author entities in the absence of unique identifiers.  Email addresses and author institutions can be used to provide information on the likely identity of an author name. However these pieces of information are often noisy and inaccurate, vulnerable to change when the author moves to a different institution or department.

I will describe a nonparametric Bayesian generative model based on the hierarchical Dirichlet process which models author entities, topics and research groups in a corpus of documents. The research groups couple together the authors by utilising coauthor information and the topics are used to differentiate the authors further. I will then describe results performed on citation databases including CiteSeer, Rexa and arXiv and compare results with the parametric Author-Topic model.


24/11/09  11:00 - 12:00

ANC workshop talks: Lawrence York and Roger Zhao, Chair: David Sterratt

Lawrence York - Encoding of stimulus pairs in higher visual areas

How single stimuli are encoded in the visual system is well known, but in typical scenes, multiple objects are present. Data shows that the response to a stimulus pair more closely matches the maximum of, rather than the sum of, the responses to the stimuli presented singly, i.e. the responses interact nonlinearly. I shall talk about encoding stimulus pairs - in particular, how non-linear encoding can increase coding capacity.

Roger Zhao - High-level and Low-level Aftereffects: Evidence for Similar Cortical Mechanisms?

The neural mechanisms for low-level processing of images have been studied in detail using animal models, focusing on early visual areas, but much less is known about the neural basis of high-level perception. An important issue is whether and how we can apply lessons learned from low-level studies (e.g. of how neurons in V1 respond to oriented line segments) to understanding high-level perception (e.g. of human faces). Visual aftereffects appear to offer a useful link -- visual perception changes systematically based on recent experience, leading to both tilt aftereffects and facial identity aftereffects.

Modelling results based on the architecture of early visual areas predict that these two effects should be very similar. However, previous psychophysical tests on humans have suggested that face aftereffects are ever-increasing as faces diverge from the population average, while tilt aftereffects are strongest around a limited range of orientations. Here we test using a much larger range of stimuli than had been used in previous face aftereffect studies, and find very similar results between FAE and TAE. The shape of the curves suggests that face-selective neurons have very broad tuning, but are otherwise similar to orientation-selective neurons. These results suggest that adaptation, and thus probably much of the organization and representation, are similar between low-level and high-level visual perception.

This talk will be similar to my DTC Day talk in September, but with new modelling results and additional experimental controls undertaken based on feedback from the earlier version.

03/11/09  11:00 - 12:00

ANC Workshop talks: Tim Lukins and Nicolas Heess, Chair: Judith Law

Tim Lukins:

"Cloudy Vision: Demystifying On-Demand Scalable Image Processing."

Perhaps the most bandied phrase of the moment is "The Cloud". But what does it actually mean, what benefits does it have, and how can it be used for processing scientific data? How easy is it to put computational/memory intensive algorithms to work on ever larger data-sets? Especially if our data happens to be image or video based. In this practically focused talk I will present a brief introduction to some of the new technologies associated with cloud based computing. I will describe the unique challenges presented by video/images, and show how code can be written and deployed to process and analyse such data efficiently. I will finally comment on further trends and solutions that are starting to appear for web-services providing advanced computer vision and machine learning.

Nicolas Heess:

"Masked RBMs for factoring appearance and shape"

Restricted Boltzmann Machines (RBMs) are capable of learning the structure of highly complex data without supervision. However, when applied to images, RBMs have difficulty representing occlusion boundaries since these represent a sharp transition from one set of image statistics to another.  I'll briefly describe the Masked RBM (proposed by N. Le Roux & J. Winn) which allows occlusion boundaries to be modelled by factoring out the appearance of an image region from its shape, representing each with a separate RBM.  I'll then discuss how to model shape in this framework and demonstrate how incorporating a notion of "occlusion" into the model makes it possible to learn a more efficient representation of shapes and also to perform inference with respect to the relative depths of image regions.  I'll discuss a version of the model for image patches and, if there is time, its extension to full images.

This is work that I did with Nicolas Le Roux and John Winn during my internship and MSR.

20/10/09  11:00 - 12:00

ANC Workshop talks: Charles Sutton, Chair: Chris Williams

Learning and Inference in Networks of Queues

As computer systems have become more complex, in part because of the use of both software libraries and parallelism, it has become increasingly difficult to (a) predict the performance of the system under new conditions, and (b) to understand the causes of poor performance of the system.  Both of these problems are ideal candidates for a machine learning approach, because probabilistic modeling provides a natural framework to combine prior knowldege of the system with mesaurements of its performance.

In this talk, I present a novel graphical modeling viewpoint on queueing models, which allows them to be used for inference about past system behavior and learning from incomplete data. The idea is to measure a small set of arrival and departure times from the system, treating the times that were not measured as missing data. The posterior distribution over missing data and parameters can then be sampled using Markov chain Monte Carlo techniques. Developing a sampler in this case is significantly more challenging than for standard graphical models, because of the complex deterministic dependencies that arise in queueing models.  To my knowledge, this is the first example of probabilistic inference in a queueing model.

Finally, I will discuss several promising avenues for future work, including model selection problems for computer systems, and generative models of computer programs more generally.

06/10/09  11:00 - 12:00

ANC Workshop talks: Jim Bednar and Matthew Down, Chair: Cian O'Donnell

James Bednar:

"Beyond V1"

The primary visual cortex (V1) in mammals has been the focus of intense reserach for more than 50 years, and we now have some well established ideas of how neurons respond to standard test patterns in anaesthetised animals.  There is much to do before we can be said to have understood what V1 does, but for further progress it is necessary to consider V1 in the context of the larger nervous system, organism, and ecological niche in which it is embedded.  In this talk, I will briefly summarize my understanding of V1 based on previous experimental and computational work, and then outline a little of this larger context, in the hopes that members of the audience can help point out the biggest gaps in my knowledge and give ideas for directions of future research.  Essentially, this will be a talk about what I do not know, and so I hope for some feedback, discussion, and interesting ideas.

Matthew Down:

"Modeling axon pathfinding at the mouse optic chiasm."

Healthy pre-natal development of the visual system requires that the projection of retinal ganglion cell axons from the retina follow distinct trajectories to their targets in the thalamus and superior colliculus. The precise nature of these trajectories at the mouse optic chiasm and their mathematical relation to gradients of molecular guidance cues is the focus of my PhD. My talk will be an update on progress so far.

22/09/09  11:00 - 12:00

ANC Workshop talks: Annie Huo & Jakub Piatkowski, Chair: Tim Lukins

Jakub Piatkowski:
'Multiple-subjects connectivity-based parcellation using hierarchical Dirichlet process mixture models' by Jbabdi et al

We propose a hierarchical infinite mixture model approach to address  two issues in connectivity-based parcellations: (i) choosing the number of clusters, and (ii) combining data from different subjects.  In a Bayesian setting, we model voxel-wise anatomical connectivity profiles as an infinite mixture of multivariate Gaussian distributions, with a Dirichlet process prior on the cluster parameters. This type of prior allows us to conveniently model the number of clusters and estimate its posterior distribution directly from the data. An important benefit of using Bayesian modelling is the extension to multiple subjects clustering via a hierarchical mixture of Dirichlet processes. Data from different subjects are used to infer on class parameters and the number of classes at individual and group level. Such a method accounts for inter-subject variability, while still benefiting from combining different subjects data to yield more robust estimates of the individual clusterings.


Annie Huo:

"Adaptive Map Alignment in the Superior Colliculus of Barn Owl: Model and Neuromorphic Implementation "

There are many kinds of brain mapping. Here we will discuss sensory map alignment in Superior Colliculus.

Neuroscience has always been the source of inspiration for researchers in artificial intelligence. Here, we describe how a specific bio-inspired model in robot and neuromorphic system. This implementation improves the system's ability to adapt the visual and auditory information integration process to environmental changes. The inspiration comes from the barn owl, a nocturnal predator which can locate its prey precisely through use of its strong visual and auditory localization system. A juvenile barn owl can adapt its visual and auditory map alignment when its visual system is disrupted by prism wearing. We explored a possible brain development mechanism of this adaptation and show its viability by hardware implementation. This can provide new insights enabling further bridges to be built between neuroscience and engineering. 

08/09/09  11:00 - 12:00

ANC Workshop talks: Andrew Pocklington and Mike Dewar, Chair: Matthias Hennig

Mike Dewar

An overview of dynamic spatiotemporal models

Spatiotemporal systems constitute a huge class of system, the study of which has led to a wide set of dynamic models. I will give a general overview of these models, some example applications and data sets, and highlight some common issues that hamper widespread adoption of these methods. I'll also introduce some of the topics that will be dealt with in next month's PASCAL workshop on spatiotemporal modelling.

01/09/09  11:00 - 12:00

ANC Workshop talks: Amos Storkey and Alex Mantzaris, Chair: Mike Dewar

Alex Mantzaris

We present an improved phylogenetic factorial hidden Markov model (FHMM) for detecting two types of mosaic structures in DNA sequence alignments, related to (1) recombination and (2) rate heterogeneity. The focus of the present work is on improving the modelling of the latter aspect. Earlier papers have modelled different degrees of rate heterogeneity with separate hidden states of the FHMM. This approach fails to appreciate the intrinsic difference between two types of rate heterogeneity: long-range regional effects, which are potentially related to differences in the selective pressure, and the short-term periodic patterns within the codons, which merely capture the signature of the genetic code. We propose an improved model that explicitly distinguishes between these two effects, and we assess its performance on a set of simulated DNA sequence alignments.

Amos Storkey

Top Challenges in Machine Learning

I will discuss what I think are some of the top medium-to-long term challenges in machine learning, why they are (a little bit) important.

25/08/09  11:00 - 12:00

ANC Workshop talks: Adrianna Teriakidis and Judith Kirkwood-Law, Chair: Adam Chai

Judith Kirkwood-Law

Topographic organization of mouse visual cortex

I have talked several times now about modeling the development of maps in mouse visual cortex, but in this talk I will be focussing on the 2-photon data on which some of my models have been based. I will show some new unpublished 2-photon imaging data of combined retinotopy and orientation maps in mouse from Tom Mrsic-Flogel (UCL). In particular I will talk about how the preference for retinal space at the level of individual cells in mouse V1 appears to be scattered. I will discuss a number of considerations which arise when trying to quantify this scatter and also whether this scatter has an influence on the topographic organization of other features in V1.

Adrianna Teriakidis

I will talk about my evidence that synapse elimination can happen in the absence of competition, make a suggestion for what might be driving this and show how I tweaked a model to test this hypothesis. I will also give a brief update on my search for converging sibling axons in developing motor neurons.

18/08/09  11:00 - 12:00

ANC Workshop talks: Douglas Armstrong and Graeme Phillipson, Chair: Seymour Knowles-Barley

Douglas Armstrong

Modelling Schizophrenia in flies.

Over the past 12 months we have undertaken a couple of pilot studies to examine what happens to flies when we express human proteins liked to Schizophrenia in their brains. Given the molecular nature of these proteins and what they interact with we hypothesised that the most likely phenotype would be in memory. Our progress on this will be presented along with preliminary results from the first two candidate genes PDE4B1 and DISC1.

Graeme Phillipson

Global vertical disparity influences stereo correspondence.

Stereovision requires the brain to find matches between image features in two retinas.  The correct match lies on an epipolar line determined by the orientation of both eyes.  Thus, information about epipolar geometry could potentially simplify stereo correspondence by indicating where matches are likely to be found.  We investigated this by asking subjects to detect a disparity-defined disc presented at 10deg eccentricity in a random quadrant.  Task difficulty was altered by decreasing the interocular correlation, making stereo correspondence harder due to the increasing number of dots with no match in the other eye.  We find that adding on-screen vertical disparities which simulate infinite viewing distance can actually increase task performance.  Thus, human stereo correspondence may concentrate its search near the epipolar lines appropriate to infinity, even when this is inconsistent with the oculomotor system.  However, an inconsistent vertical disparity field, where the periphery (>20deg) indicates the correct 30cm viewing distance while the region <20deg indicates infinite viewing distance, has an even more detrimental effect.  Thus, the brain may not simply search a fixed zone on the retina, but may adjust its search according to the epipolar geometry indicated by the global vertical disparity field. 


11/08/09  11:00 - 12:00

ANC Workshop talks: Mark van Rossum, Chair: Lawrence York

Mark van Rossum

Transmission of population coded information.

As neural activity is transmitted through the nervous system, neuronal noise is believed to degrade the encoded information and limit performance. It therefore is important to know how information loss can be prevented. Using a principled approach we show that information loss in layered networks depends strongly on the connectivity and that layers are best connected with a center-surround structure with a spatial extent matching the neurons' tuning curves. With this connectivity, population coded information can propagate almost without loss. The eventual degradation of information occurs through the buildup of correlations. This suggests that the center-surround architectures, ubiquitous in the
nervous system, provide a natural way to communicate population coded information.

04/08/09  11:00 - 12:00

ANC workshop talks: David Willshaw and Edwin Bonilla, Chair: Tim Lukins

Edwin Bonilla

MilePost GCC: A machine-learning based adaptive optimising compiler

Tuning compiler optimisations for rapidly evolving computer architectures makes porting and extending an optimising compiler for a new platform extremely challenging. In the MilePost (Machine Learning for Embedded Programs Optimisation) project we have developed compiler technology that can automatically learn how to best optimise programs for re-configurable heterogeneous embedded processors based on machine learning techniques.
In this talk I will discuss the past, present and future of machine learning approaches to adaptive program optimisation and will briefly describe MilePost GCC: a machine-learning based adaptive optimising compiler.

David Willshaw

Order out of disorder

In this talk I describe my recent work on applying my method of assessing the order in somatotopic maps of nerve connections. I have used the data published in Cang et al, (J Neuroscience, Vol 28, 11015-11023 (2008)), who employed the fourier-based intrinsic imaging approach to construct maps of the direct projection of retina onto superior colliculus in mouse.

Here I discuss the retinocollicular maps in mice in which some of the genes that control either molecular signalling or electrical signalling have been knocked out. In normal animals there is an ordered map of the retina onto the superior colliculus. In the electrical signalling knockouts, fairly good maps are still formed; in the molecular signalling knockouts, at first glance the maps look disordered but at a finer level they display strong and unusual types of order, as I will describe; an explanatory model for how such maps are formed is urgently needed.

28/07/09  11:00 - 12:00

ANC workshop talks: Seymour Knowles-Barley, Chair: Jakub Piatkowski

DIADEM Challenge - Hackathon Progress

In mid July we held a week long hackathon to investigate the DIADEM Neuronal Reconstruction challenge (  Several ideas for tracing methods were explored, including steerable filters, auto-focus, 3d receptive fields, betweenness, and a snake.  In this workshop we will present some of the ideas, algorithms and results.  Ample time will be left for group discussion and suggestions.

21/07/09  11:00 - 12:00

ANC workshop talks: Chris Williams and Nigel Goddard, Chair: Rob Kyle

Chris Williams 

Transfer Learning.

14/07/09  11:00 - 12:00

ANC workshop talks: David Sterratt and Lysimachos Zografos, Chair: Jan Antolik

Lysimachos Zografos

The Molecular Machinery of Cognition: Comparative Approaches and the Evolution of the Postsynaptic Proteome

Over the last year we've integrated proteomics data with various protein/gene annotations into descriptive protein-protein interaction (PPI) based models of the postsynaptic density (PSD) of the mouse and fruitfly. We've used various statistical and network science approaches to analyze the individual models, their topology, architecture and implication in behavioral phenotypes or disease. As a next step, we thought it would be interesting to compare PSDs of different organisms the models and look for the underlying principles of PSD proteome evolution. Evolutionary analysis based on sequence provides some insight, but not any linage specific information, so we decided to approach the issue via direct comparison of the two models.

In this talk I'm going to briefly summarize the state of the art of the postsynaptic proteome PPI models as well as the evolutionary analysis based on genomics sequences. Then I will give an overview of comparative interactomics methods and describe their application to my ongoing computational and proteomics work.

David Sterratt

The DIADEM Neuronal Reconstruction Challenge

Following our discussion at the ANC Day in May, Adrianna Teriakidis has helped to put together a team of DTC students and postdocs to tackle the DIADEM Neuronal Reconstruction challenge ( The aim of the competition is to derive the structures of neurons within images generated by conventional, confocal or multiphoton microscopy.

The group has been holding weekly meetings, and from 13th-17th July will be holding a "hackathon" in 2.53. I will outline the challenge and report on the progress made by the group so far.

30/06/09  11:00 - 12:00

ANC workshop talks: John Quinn and Catherine Ojo, Chair: Matthew Down

John Quinn

Cops, Crops and Mobile Phones: Machine Learning in Africa

In the developing world a great deal of data is collected which is rarely fully exploited by AI researchers. I will introduce three applications of machine learning that we are working on in Makerere University (Uganda): prediction and surveillance of disease outbreaks using varied sources of data, robust traffic monitoring from CCTV images and crop disease diagnosis with computer vision.

16/06/09  11:00 - 12:00

ANC workshop talks: Andrew Dai and Iain Harlow, Chair: Graeme Phillipson

Iain Harlow

Controversial fMRI: What exactly can we trust?

In a recent, widely publicised meta-analysis (Vul et al. 2009) Edward Vul and colleagues question the validity of a number of recent fMRI studies of emotion, personality and social cognition - including several from high-impact journals such as Science and Nature. They claim that results from a “disturbingly large, and quite prominent, segment of fMRI research… should not be believed”. But what exactly is this based on, and to what extent do their concerns generalise to the wider field of fMRI research?

In this talk I will outline the arguments made in the original paper, and why it has generated such controversy. I will also outline some of the responses from other researchers in the field, and briefly discuss the effect these issues have on interpreting reported fMRI data, from the expert reviewing journal submissions to the interested layperson.

Andrew Dai

Entity resolution or record linkage is a difficult problem that is often the first step in data mining to reduce duplication in a dataset. When applied to authors and citation databases, this becomes a problem of discovering real-world author entities in the absence of unique identifiers. Email addresses and institution information are often unavailable in citations and when they are available, they can change as authors move from place to place. In any given citation database the real number of author entities is also often unknown. In my talk I will introduce my nonparametric Bayesian model that is based on and extends the hierarchical Dirichlet process and latent Dirichlet allocation to tackle this problem. I use information about co-authors of a paper as well as their inferred associated topics to jointly infer the author entities from citations and abstracts. I will also present my latest results on the real-world arXiv KDD and CiteSeer citation datasets.

09/06/09  11:00 - 12:00

ANC workshop talks: Cian O'Donnell and Matthias Hennig, Chair: Alex Mantzaris

Cian O'Donnell

Dendritic spines as devices for synaptic metaplasticity

Most excitatory connections between neurons in the brain terminate on small protrusions called dendritic spines. Although the exact function of these spines is unclear, experimental studies have shown that they can act as localised calcium compartments. These calcium signals are widely believed to trigger synaptic plasticity - the main substrate of learning and memory in the brain. I'll present some recent modelling work testing the idea that by changing their shape and size, dendritic spines can dynamically regulate synaptic learning rules. Our results suggest that different configurations could force synaptic weights to become stable, unstable or even bistable. This can be thought of as 'metaplasticity'.

Matthias Hennig

Computational models of neural development and homeostasis

Homeostatic processes are an important mechanism for stabilising electrical activity in neural circuits. Experimental work has not only shown that they critically contribute to nervous system development, but also suggests that they are also responsible for compensating for disturbances encountered by the mature brain. Despite growing evidence that homeostatic processes constitute a complicated regulatory network of molecular pathways inside each neuron, current models are typically based on simple feedback mechanisms controlling neural activity. This substantial simplification makes it currently very difficult to relate molecular pathways to function, and is also likely to ignore important aspects of homeostasis in neurons altogether. In this talk I will give an outline of a planned research project to address these problems by means of computational and mathematical modelling. I will argue that an approach combining methods from computational neuroscience and systems biology, and the analysis and incorporation of experimental data is the way forward, which will enable us to study these and related problems that are experimentally still extremely challenging.

21/04/09  11:00 - 12:00

ANC workshop talks: Matthew Down, Chair: Xavier Oliver Duocastella

Matthew Down

"Modeling axon pathfinding at the mouse optic chiasm."

Healthy pre-natal development of the visual system requires that the projection of retinal ganglion cell axons from the retina follow distinct trajectories to their targets in the thalamus and superior colliculus. The precise nature of these trajectories at the mouse optic chiasm and their mathematical relation to gradients of molecular guidance cues is the focus of my PhD. My talk will be an update on progress so far.

14/04/09  11:00 - 12:00

ANC workshop talks: Lawrence York and Jim Bednar, Chair: Mike Dewar

Lawrence York: Processing multiple stimuli with different latencies in the visual cortex

Recent experimental evidence indicates that some neurons in the visual system responding to two stimuli suppress the response to the later stimuli over short timescales. I will discuss several models which could account for this phenomenon, whilst presenting the surrounding context of experimental work in this area.

James A. Bednar: Modeling the Emergence of Whisker Direction Maps in Rat Barrel Cortex

By measuring response to mechanically stimulated rat whiskers, Andermann and Moore (Nat. Neurosci 9:543-551, 2006) demonstrated that barrel fields in rat primary somatosensory cortex (S1) contain a pinwheel map of whisker motion directions. Because this map is reminiscent of topographic organization for visual direction in primary visual cortex (V1) of higher mammals, we asked whether the S1 pinwheels could be explained by an input-driven developmental process as is often suggested for V1. We developed a computational model to capture how whisker stimuli project to supragranular S1, and simulate lateral cortical interactions using the LISSOM self-organizing algorithm. In the model, if whisker deflection direction correlates with the pattern of whiskers that is deflected, a somatotopically
aligned map of direction emerges for each whisker in S1. This result suggests that similar intracortical mechanisms guide the development of primate V1 and rat S1, and we propose specific experiments for testing this hypothesis by manipulating whiskers during early development.

This is a greatly revised version of work first presented to ANC in Feb 2008, with much simpler and clearer results.

Joint work with Stuart P. Wilson (primary author), Ben Mitchinson, and Tony J. Prescott from U. Sheffield, and with Judith S. Law from U. Edinburgh.

31/03/09  11:00 - 12:00

ANC workshop talks: Douglas Armstrong and Rob Kyle, Chair: Juan Huo

Rob Kyle

Understanding the brain in terms of the self: Neuroscience beyond reinforcement learning and neuromodulation.

In recent years it has become popular to look at the function of the brain as a kind of reinforcement learning circuit, with neuromodulators such as dopamine, serotonin, and acetylcholine acting as parameters for the system, signalling reward and expectation. In fact a recent speaker went so far as to say that this model is "the only theory of the brain we have in neuroscience"....

Is this a fair statement, and should we expect the research in this area to continue to grow?

In this talk I aim to look towards where such a model of cognition might lead us, and in particular examine how such a model of perception and action can be integrated with an organism's evolutionary critical need for self-preservation. I will describe how the model of reinforcement learning and neuromodulation implies the existence of a heirarchy of modulators with their nexus located in the interaction between the immune, endocrince, and nervous system. I will introduce a potential model for understanding this interaction and time permitting, describe biological evidence which might be used to support this view.....

Douglas Armstrong

Towards a virtual fly brain

Models of the brain that simulate sensory input, behavioural output and information processing in a biologically plausible manner pose significant challenges to both Computer Science and Biology. We have investigated strategies that could be used to create a model of the insect brain, specifically that of Drosophila melanogaster which is very widely used in laboratory research. The scale of the problem is an order of magnitude above the most complex of the current simulation projects and it is further constrained by the relative sparsity of available electrophysiological recordings from the fly nervous system. However, fly brain research at the anatomical and behavioural level offers some interesting opportunities that could be exploited to create a functional simulation. We propose to exploit these strengths of Drosophila CNS research to focus on a functional model that maps biologically plausible network architecture onto phenotypic data from neuronal inhibition and stimulation studies, leaving aside biophysical modelling of individual neuronal activity for future models until more data is available.

24/03/09  11:00 - 12:00

ANC workshop talks: Judith Law and Mike Dewar, Chair: Lysimachos Zografos

Judith Law

I will be presenting the final part of my PhD thesis where I have been exploring the differences between rodents and higher mammals that might lead to the lack of topographic organization for orientation in rodent species. Analysis of retinotopy data from two-photon calcium imaging in mouse (provided by Tom Mrsic Flogel, University College London) shows that at this resolution the retinotopic map in mouse is highly disordered. Parameterized models of mouse and cat visual cortex are used to investigate a number of possibilites for how disorder in retinotopy and orientation preference may arise in the mouse and why this might differ between species. These results suggest that species with and without topographic maps implement similar visual algorithms differing only in the values of some key parameters, rather than having fundamental differences in architecture. However a number of targeted experimental studies will be required in order to distingish the key parameter differences.

Mike Dewar

Inferring the behaviour of Drosophila

Over the past ten months I've been involved in a proof-of-concept project whose aim was to demonstrate the viability of automatically inferring the behaviour of animals from video data. In this talk I am going to present the machine learning associated with this project, using, as a case study, courtship behaviour in Drosophila. I'll describe a static approach to the problem, followed by subsequent attempts to use dynamic models to properly characterise their behaviour. I'll conclude by proposing a simple modelling framework that I think encapsulates most automated behaviour analysis, along with some ideas for future work.

10/03/09  11:00 - 12:00

ANC workshop talks: Alex Mantzaris and Andrew Pocklington, Chair: Edwin Bonilla

Alex Mantzaris

Addressing potential fallacies in Bayesian methods, for detecting recombination.

Various probabilistic models based on change-point processes and , HMMs have recently been proposed to detect tree topology, changes in phylogeny resulting from recombination. When treating the inference problem in the Bayesian, framework, the marginal posterior probability of the tree, topology at a given site in the DNA sequence alignment has to, be computed, which requires the solution of an integral over , the space of branch lengths and nucleotide substitution parameters. To render this integration analytically tractable, several authors, have recently made the assumption that the vectors of branch lengths , of the phylogenetic tree are independent among sites. While this , approximation reduces the computational complexity considerably , we demonstrate with a Bayesian hypothesis test, that it leads to the systematic prediction of spurious topology , changes in the Felsenstein zone, that is, the area in the branch , lengths configuration space where maximum parsimony consistently , infers the wrong topology due to long-branch attraction., We discuss how the problem is related to the, no-common-mechanism model of Tuffley and Steel, and how it , results from failing to appreciate the difference between, structural and incidental model parameters., We conclude by proposing a novel phylogenetic factorial HMM, that addresses these shortcomings. This is joint work with Dirk Husmeier.

Andrew Pocklington

Molecular models of synapse function, evolution and role in disease.

There are currently no models of synaptic signalling that reflect its molecular complexity. I will briefly review work from the last few years in which we've sketched out a rough picture of the functional organisation and evolutionary development of the post-synaptic signalling machinery. I will end with a quick summary of some recent results linking this work to the study of schizophrenia.

03/03/09  11:00 - 12:00

ANC workshop talks: Simeon Bamford and James Withers, Chair: Wan-Yu Hung

James Withers

Multi-disciplinary approach to brain MRI segmentation

Brain images produced by structural MRI scans often possess variable contrast between the main tissue types, intra-class intensity inhomogeneity, poor resolution and relatively high amounts of noise. Accurate segmentation is further complicated by partial volume structures, thin and unique anatomy, as well as inter- and intra-rater differences between human experts producing the ground truth. In this talk I will compare the results of my method on a challenging poor quality dataset of aged patients against a leading automated brain segmentation tool (FSL FAST). Furthermore, past and proposed work covering human computer interaction, graphical visualisation, machine learning and bioinformatics fields will be presented.

24/02/09  11:00 - 12:00

ANC workshop talks: Matthias Hennig and Adrianna Teriakidis, Chair: Cian O'Donnell

Matthias Hennig

Sound source localisation in Kcna1-null mice

In this talk, I will present results from a little side-project I did last year in collaboration with Conny Kopp-Scheinflug in Leipzig. We investigated responses of neurons in different auditory brainstem nuclei of mice that are involved in sound source localisation. Activity in normal mice was compared to responses in mice where a particular low-voltage activated potassium channel subunit (Kcna1) had been genetically removed. These experiments showed that this manipulation affected in particular the ability of neurons to reliably generate precisely timed spike trains at high rates, and that sound source localisation was impaired as a result. We developed a simple computational model to analyse these effects. These simulations not only reproduce the changes observed in the knockout mice, but also shed new light on how sound source localisation is achieved by the brain.

Adrianna Teriakidis

From the beginning of my PhD I have been trying various approaches to get data about neonatal motor neurons, and specifically trying to find out how many synapses individual motor neurons make.  The problem is that the resolution of confocal microscopy isn't enough to be able to distinguish between the axons when they come close together.  Last term I spent three months in Jeff Lichtman's lab trying out a technique they are developing for increasing the resolution in the z-direction.  This involves physically sectioning the sample, imaging the sections and then re-aligning them.  I will talk about the technique, the problems I encountered with it and speculate as to why it hasn't worked.  

17/02/09  11:00 - 12:00

ANC workshop talks: Jan Antolik and Graeme Phillipson, Chair: David Sterratt

Jan Antolik

While there are a number of computational models for how specific aspects of primary visual cortex can develop, each of these focuses on explaining a narrow range of observations. Each of these observations can be explained by a wide, diverse range of possible models, which makes it very difficult to determine which model is correct (if any). In this work I will present our work on developing a model of V1 development that is constrained by experimental data of widely different sources and types, to show that (a) it is difficult to devise a model compatible with all of this evidence, and (b) it is nonetheless possible. In particular, relative to existing models of orientation map development, I will show how adding homestatic regulation of responses to inputs can explain the stability of maps over the course of development, despite very large variability in the type and amount of activity-based input. I will also show that the individual cells in these maps have realistic orientation tuning curves and realistic contrast response, thus replicating features previously only demonstrated in single-cell models. The overall goal is to account for observations across a wide range of spatial and temporal scales with a single framework driven by Hebbian learning of firing rates in simplified neural models.

Does stereo correspondence account for eye position?
Graeme Phillipson & Jenny Read School of Biology, University of Newcastle

In order to fuse the two eyes' images into one, the brain has to work out which point in the left retina is viewing the same point in space as a given point in the right retina -- the stereo correspondence problem. At first sight this seems like a two-dimensional problem -- the correct match could be anywhere in the lef retina. But in fact, if eye position is known, geometry reduces this to one-dimension: the correct match lies somewhere on an "epipolar line" on the left retina. Artificial stereo systems use this constraint, updating the epipolar lines each time the cameras move, in order to make the search for correspondences efficient. But surprisingly, there is currently no evidence that the brain takes account of eye position in solving the stereo correspondence problem. I will describe a set of experiments designed to reveal if the brain takes account of the movement of epipolar lines which occur when the eyes verge at different distances.

10/02/09  11:00 - 12:00

ANC workshop talks: Alex Mantzaris, Chair: Andrew Dai

27/01/09  11:00 - 12:00

ANC workshop talks: David Willshaw and Iain Harlow, Chair: Nicolas Heess

David Willshaw

High precision visual maps and their interpretation.

I'm going to describe a relatively new method for constructing visuotopic maps in the vertebrate nervous system.  This was developed in Michael Stryker's lab at UCSF and involves fourier-based intrinsic imaging.  I will then describe the application of my method for measuring map precision to imaging data of this type from Jianhua Cang's lab (J.Neuroscience, Vol 28, 11015-11023  (2008).  I use this method to explore the difference in precision between normal maps in mouse and the maps disturbed by knockout of the Beta2 subunit of the acetylcholine receptor, which affects the pattern of electrical.


Iain Harlow

Can familiarity support associative memory?

Understanding how episodic (autobiographical) memories are stored, accessed and related to each other is a crucial aspect of understanding consciousness: every person is a product of their experiences.

It is currently thought episodic memory for individual stimuli can be supported by two functionally and neurally separable processes. Familiarity is a sense of having encountered the stimulus before (I know that face...); while Recollection involves the conscious retrieval of details about the episode (oh, it's the barman who served me yesterday).

How these processes support associative memory (links between separate items or memories) is controversial: some argue that recollection is necessary, while more recently it has been suggested that under certain circumstances familiarity can also play a role.

In this talk I will present some of my own data suggesting familiarity may indeed support associative memory. If there is time, I will also discuss some issues involved with modelling and measuring familiarity and recollection, and some future work intended to address these issues.


20/01/09  11:00 - 12:00

ANC workshop talks: Xavier Oliver Duocastella and Chris Williams, Chair: Iain Harlow

Xavier Oliver Duocastella

The development of the thalamocortical system

The fibers connecting cerebral cortex and thalamus form a very important tract in the brain. This talk will aim at introducing the thalamocortical system and the tract connecting it reciprocally, with a special interest in its development in the embryonic mouse. The first attempts at doing a 3D model of the developing thalamus will also be discussed.


Chris Williams

Top-down and Bottom-up Processing in Probabilistic Models of Visual Perception

Many models of visual processing tasks are feed-forward. However, locally visual patterns are often highly ambiguous and subject to multiple interpretations. Contextual influences (top-down or lateral) are often necessary in order to disambiguate the interpretation.

In this talk I will first discuss unsupervised learning of probabilistic models as specifying (or constraining) normative models of the visual system. I will discuss examples of contextual effects in visual processing, and probabilistic models that incorporate contextual effects (e.g. MRFs, tree-structured belief networks directed hierarchies of MRFs).

13/01/09  11:00 - 12:00

ANC workshop talks: Nicolas Heess and Chris Palmer, Chair: Matthew Down

Nicolas Heess

Learning generative texture models using extended Fields-of-Experts (work with Chris Williams)

In the context of learning hierarchical generative image models I have been investigating the ability of the Fields-of-Experts (FoE) model (Roth & Black, 2005) to model different types of textures. Although the "naive" FoE model learned with contrastive divergence (Hinton 2000) as proposed by Roth & Black does not learn good generative models, I will demonstrate that two simple modifications, (a) to the learning procedure and (b) to the structure of the model can substantially improve its generative power. The extended model produces good results for a range of natural and synthetic textures in 1D and 2D. I also compare these results to the performance of the simpler Gaussian FoE. I'll give some intuitive explanation why the extended model works better than the standard FoE, why especially the Gaussian model is unsuitable for the task at hand, and what one might learn from this for image modeling in general.

Chris Palmer

Topographic and laminar models for the development and organisation of spatial frequency and orientation in V1

Speculation has existed for some time over the possible organisation of spatial frequency (SF) preference in mammalian visual cortex (V1). A consensus appears to be emerging that in the superficial layers SF is mapped continuously across the cortical surface. However, other evidence suggests that SF may differ systematically with cortical depth, at least in layer 4, where the M and P pathway afferents terminate in different sublaminae. It is not yet clear whether the topographic organization for SF observed in the superficial layers is maintained throughout the input layers as well, or whether there is a switch from a laminar to a topographic organization along the vertical dimension in V1. We present results from two alternative self-organising computational models of V1 that receive natural image inputs through multiple SF channels in the LGN, differing in whether they develop laminar or topographic organisation in layer 4. Both models lead to topographic organization for OR and SF preference in upper layers, consistent with current experimental evidence. The results suggest that in either case separate sub-populations of neurons are required to obtain a wide range of SF preference from Hebbian learning of natural images. We demonstrate that the same principle of representation can arise within different neural architectures and ultimately achieve the same outcome - this may parallel the differences in feature preference representation which have been observed in different species.

06/01/09  11:00 - 12:00

ANC workshop talks: Adam Chai, Chair: Chris Palmer

Adam Chai

Approximate probability regression --- GP binary classification without (much) pain

Binary classification is an important application of Gaussian Process in machine learning. Unlike the regression case, the posterior of the latent process cannot be obtained analytically, and has to be approximated. The Laplace method approximate the posterior by a Gaussian fit to the Hessian of the mode, while the EP method approximate it by a Gaussian matching the moments at the marginals. Both procedures involve iteration. In this note, we propose an approximate probability regression model that directly fit a multivariate Gaussian by matching the prior up to the second co-moments; no iteration is required. This method is restricted to the common case where the prior mean of the latent process is a constant zero.

16/12/08  11:00 - 12:00

ANC workshop talks: Seymour Knowles-Barley, Chair: Liana Romaniuk

Seymour Knowles-Barley

How to Scan a Fruit Fly Brain

The post synaptic density is a dense region of proteins located at the synaptic cleft between two neurons. This region plays a major role in synaptic plasticity. Exactly what proteins make up this region and how they interact is not well understood. In this workshop I will present a website databasing the 3d expression patterns of proteins in the fruit fly brain and results from Electron Microscopy scans attempting to image post synaptic density proteins at 1nm resolution.

09/12/08  11:00 - 12:00

ANC workshop talk: Mark van Rossum and David Sterratt, Chair: Simeon Bamford

Mark van Rossum

Modeling synaptic plasticity: Longevity, tagging and capture

In recent years we have seen a number of experimental results that refine and challenge our understanding of synaptic plasticity. Some of the questions that these experiments address are: Is synaptic plasticity between different synapses independent? How do reward and plasticity interact? What determines the longevity of synaptic plasticity. The functional consequences of the experimental findings is not always clear. We present a computational state-based model that integrates these finding. The model will form the basis for more realistic network simulations.

David Sterratt

Inferring spherical coordinates from flattened retinae

In retrograde tracing experiments to determine the mapping of connections from the retina to the superior colliculus in mice, a small blob of dye is injected in the superior colliculus and allowed to diffuse retrogradely down the axons of retinal ganglion cells to their cell bodies in the retina. The retina is dissected and flattened, and
the pattern of dye in cell bodies can be seen in the flattened retina. In the process of flattening the retina, a number incisions are made, and the pattern of dye can cut across incisions, complicating analysis. One way of simplifying the analysis would be to infer the position of the cell bodies in the spherical coordinate system of the intact retina. his talk is about my attempts to do this so far.

02/12/08  11:00 - 12:00

ANC workshop talks: Cian O'Donnell and Andrew Dai, Chair: Jesus Cortes

Cian O'Donnell

Biophysics matters: Hodgkin-Huxley and ion channel noise

Today we know that the electrical signals computed and transmitted by neurons are generated by large numbers of discrete ion channels embedded in their cell membrane. Each of these ion channels opens and closes only stochastically. However, many mathematical models of neurons (like Hodgkin-Huxley) model their large-scale behaviour with deterministic differential equations. Although this approach can explain many experimental findings, it omits the intrinsic noise from stochastic channel gating. In such non-linear systems, this noise can often have surprising consequences. We represent individual channels as Markov models and find that their population behaviour can be qualitatively different from the deterministic system, causing membrane potential noise, spontaneous action potentials and degraded spike reliability.

Andrew Dai

Author disambiguation with Bayesian nonparametrics.

One of the problems that citation databases have is how to identify and distinguish authors among a corpus of documents. Most authors don't use any kind of unique identifier when authoring papers so authors with the same name are often confused to be the same person. On the other hand, authors that have differing names may actually be the same person and the differences may arise from OCR or transcription errors or more rarely from name changes such as from marriage. The result of this is that in a large corpus of documents the true number of author identities is unknown.

I'll describe our Bayesian nonparametric model that is used to model a set of papers where the authors have names which both contain errors and may not be unique. This involves modelling the authors and document topics together in a generative model that makes no fixed assumption on the total number of authors or topics in the corpus.

25/11/08  11:00 - 12:00

ANC workshop talks: Jakub Piatkowski and Peggy Series, Chair: Rob Kyle

Peggy Series

Sensory illusions: why do they occur ?

I will review some ideas related to the origin of sensory illusions, and the current theoretical understanding of these effects, inspired from my own work and from a recent article by Odelia Schwartz, Anne Hsu and Peter Dayan, Nature Neuroscience Reviews, 2007.

Jakub Piatkowski

Measuring white matter volume (and integrity) with diffusion MRI.

In order to study white matter in health and disease, a reliable way of estimating the parameters of interest, such as tract volume or coherence, must be found. Diffusion MRI is the only non invasive tool available; it's now widely used in clinical practice and basic neuroscience.

Considerable amount of work in the recent years has focused on correct white matter fibre angle estimation and pathway tracking, which can now be done reasonably well. Tract integrity, however, is still estimated using an index (FA) derived from the earliest diffusion model (DT), even though it is well known that this estimate is not reliable. Worse still is the estimation of tract volume, for which there are no well established methods of acceptable precision.

My PhD project is aimed to fill this gap by developing a method to describe white matter tracts completely, estimating their actual physical extent as well as coherence

18/11/08  11:00 - 12:00

ANC workshop talks: Lysimachos Zografos and Amos Storkey, Chair: Tim Lukins

Amos Storkey

In many domains, modelling is naturally done in continuous time. The system is sensibly specified in terms of a differential equations. Because of uncertain inputs, contributions or hidden factors, the differential model is rarely exact and so a stochastic differential model (SDM) is the most appropriate.

In SDM's it is possible to perform filtering (i.e. forward inference) efficiently using variable step size stochastic Runge-Kutta approaches. However smoothing (the combination of forward and backward inference) is much more problematic for a number of reasons. One is the unavailability of the explicit forward transition density over finite time, the second is the need, under most schemes to reuse particles from the forward pass, hence constraining the variability in the time step for the backwards pass, and the third is the typical degeneracy problems which occur from mismatch of the forward and backward messages.

We utilise a kernel density reweighting scheme to enable the decoupling of the forward and backward passes, and use partition trees and additional parallelisations in order to provide speed enhancements. We demonstrate the approach on modelling the network interaction and hemodynamic response of simulated and real fMRI data.

This work is the subject matter of Lawrence Murray's PhD.

Lysimachos Zografos

"Modelling the Molecular Machinery Of Cognition"

The molecular processes of cognition include molecular computation during signal transduction and regulatory cascades that affect abundance and localization of molecules in the synapse. These processes are mediated by molecules residing in the post (and pre) synaptic terminals (synaptic proteome - SP). In this talk I will present a case study of modelling proteomics data from the SP, using protein interaction networks. I will also show how this models can be enriched with functional annotation and graph analysis, providing insight into how the SP is organized in modules and pathways and how the latter are involved cognition, mental disease and addiction.

11/11/08  11:00 - 12:00

ANC workshop talks: Matthew Down and Bilal Malik, Chair: Roger Zhao

Matthew Down

Modelling axon pathfinding at the developing mouse optic chiasm.

The connection between the eye and the brain is mediated by Retinal Ganglion Cells (RGC), in order for this connection to be created during development each RGC sends out a growth cone that senses extra-cellular cues and lays down the path of the axon. I shall talk about understanding the relationship between these paths and the molecular cues guiding the growth cone in terms of a particular mathematical model (Krottje and Van Ooyen 2007) giving attention to the mouse optic chiasm.

Bilal Malik

Our recent progress in characterisation of MASC complex in Drosophila .

In this talk I will talk about our recent progress in characterising the MAGUK associated signalling complex (MASC) in Drosophila. Drosophila is at an intermediate position between unicellular organisms and the multicellular vertebrates. The MAGUK seems to have undergone significant changes in its composition through the course of evolution. I will try to highlight some of these changes and also how this complex looks like in Drosophila.

04/11/08  11:00 - 12:00

ANC workshop talks: Lawrence York and John Davey, Chair: Jan Antolik

Lawrence York

I will be presenting work on the dynamics of ring attractor networks with depressing synapses, in particular some of the mathematical results regarding the stability of some network dynamics, as well as presenting a detailed examination of the intrinsically propagating activity profiles that arise in such networks, and discussing the implications for some common uses of the ring attractor model.

John Davey

As my thesis is now nearing completion, I will give an overview of my work over the last three years, identifying RNA molecules in thalamocortical axons. I will discuss the implications of these findings for cortical development. I will also discuss the job I am about to move on to, where I will be attempting to discover the gene which breaks left-right axis symmetry in bilateral animals by sequencing the transcriptome of pond snails.

21/10/08  11:10 - 12:00

ANC workshop talks: Nigel Goddard, Chair: Edwin Bonilla

Nigel Goddard

Computational Economics: two approaches

Economists study how individual preferences combine in the context of market systems to generate an allocation of resources for producing goods and services. The unit of measurement is financial – dollars, pounds, constant dollars, etc. But money is a fickle measure. Via credit creation, the financial worth of a country can increase much faster than the real wealth of the country as measured by the physical stock of goods and the physical infrastructure on which services are based. And, as we have seen recently, financial worth can decrease even faster – billions and even trillions of pounds of capital can be destroyed in a day, with no corresponding change in the physical world. Iceland provides a recent example.

Is there a better numeraire? Can we construct models of the economy based in physical reality rather than the money abstraction? The first part of this talk describes an approach to economic modeling based on energy as the numeraire, and using the system-dynamics modeling paradigm to construct a computational model that can be used to explore the implications of economic policy choices.

In the second part of the talk (if there is time!), I will introduce agent-based computational economics (ACE). Economists are still mostly mathematicians, with a strong preference for models that can be expressed and explored analytically. The concept of “equilibrium” reigns – the desirable price for a good is that reached when buyers and sellers are in equilibrium. ACE is more of an experimental approach which explores the dynamics of an economic system, without the assumption that there is necessarily any equilibrium to be reached, and without the requirement for analytical tractability. This part of the talk will be grounded in an ACE model of the European Emissions Trading Scheme, the flagship policy for the EU to reduce carbon emissions to mitigate climate change.

14/10/08  11:00 - 12:00

ANC workshop talks: Tim Lukins and Jim Bednar, Chair: Natasha Dare

Jim Bednar

Building A Comprehensive Model for Development and Function of the Visual Cortex
Vision is a complex and difficult task to understand and explain, and researchers have used a very wide range of different experimental and theoretical approaches in this endeavor. These approaches tend to have quite different assumptions, strengths, and weaknesses. Computational models of the visual cortex, in particular, have typically implemented either a proposed circuit for part of the visual cortex of the adult, assuming a very specific wiring pattern based on findings from adults, or else attempted to explain the long-term development of a visual cortex region from an initially undifferentiated starting point. Previous models of adult V1 have been able to account for many of the measured properties of V1 neurons, while not explaining how these properties arise or why neurons have those properties in particular. Previous developmental models have been able to reproduce the overall organization of specific feature maps in V1, such as orientation maps, but the neurons in the simulated maps behave quite unlike real V1 neurons, and in many cases are not even testable on actual visual stimuli because the developmental models are so abstract. I believe that the complex adult circuitry only makes sense when considering the developmental process that created it, and conversely, that the developmental process only makes sense if leading to a system that can perform behaviorally relevant visual tasks.

Accordingly, in this talk I outline a "meta-project" of coordinating a number of independent student projects in order to build the first model to explain both the development and the function of V1. That is, we hope the resulting model will be the first that starts from an initially undifferentiated state to wire itself into a collection of neurons that behave, at a first approximation, like those in V1. To do this, we are building the first developmental models with wiring consistent with V1, the first to have realistic behavior with respect to visual contrast, the first to include all of the various visual feature dimensions, and the first to include all of the major sources of connectivity that affect V1 neuron responses. The goal is to have a comprehensive explanation for why V1 is wired as it is in the adult, and how that circuitry leads to the observed behavior of the neurons during visual tasks.

Tim Lukins

Where's my mouse? - Reliable boundary based tracking of rodent shaped objects.

Many labs investigating behaviour study rodents, whose activity they record, generating large volumes of video data. Human observers then analyse the footage in order to mark-up the rodent's activity. However, humans performing such tasks are prone to inaccuracy, inconsistency, and fatigue. A more appealing solution would be to automatically track the rodent and to also label specific parts of the segmented boundary (e.g. as the head, tail, body, etc.). This is a challenging problem for Computer Vision since rodents are small, quick moving and highly deformable.

Our main assumption is that shape information alone is generally sufficient to identify physiological features and the overall orientation. I will describe our initial approach using background modelling for segmentation, followed by multi-scale curvegram analysis of boundaries, along with medial axis fitting. This engineered approach is reasonably accurate and computationally fast, but can fail on more complex videos with multiple targets and cluttered environments. Consequently, I will also present the basis for a revised attempt at a more principled approach, in the context of the current state-of-the-art regarding particle-filters and robust boundary detection.

30/09/08  11:00 - 12:00

ANC workshop talks: Douglas Armstrong and Rob Kyle, Chair: Joanna Young

Douglas Armstrong

Automatic analysis of behaviour data by supervised machine learning

Animal behaviour is the most complex of phenotypes with genetics, experience, environment and social factors all interplaying. Manual scoring of behaviour is a labour intensive and error-prone task that is not amenable to modern high-throughput methods. Here we examined the variability between experts and the possibility to learn, by example, a complex behaviour using methods from artificial intelligence. An example was provided by the analysis of the courtship ritual displayed by male Drosophila that has provided important insights into the genetics of behaviour and the development of sexually dimorphic neural circuitry. We asked a range of experienced researchers to annotate example videos using a widely published description of the behaviour. Not surprisingly we found a high degree of variability between our experts observing the same videos. Reassuringly we observed a much lower level of intra-observer variability. Detailed comparison of the expert annotation revealed two groups with systematic differences in their application of the accepted behavioural description. We tracked the position and orientation of the animals in each of the videos and subsequently used a decision-tree method to learn a set of rules to classify the behaviour. We demonstrated that the trained system can then convincingly mimic a real expert from either group.

Robert Kyle

"Does neuromodulation rather than spikes cause synaptic plasticity?"

Since Donald Hebb's 1949 postulate it has been widely assumed that the basis of memory and learning lies in persistent changes at the synapse. In it's original formulation Hebb's postulate suggested that these changes occured as a result of neurons firing action potentials, but more recent research has made it clear that the crucial factor for inducing a persistent change is calcium influx to the synapse. While pre and postsynaptic action potentials do offer a mechanism for calcium influx there are many other means by which calcium can enter the synapse, particularly through changes in conductivity and intrinsic excitability brought about by neuromodulators.

In my talk I will describe work in which I have implemented a simple model of dopamine modulation, and examined the effects this has on a calcium based model of plasticity. The question I want to ask is this: Does neuromodulation have a more significant effect on synaptic plasticity than patterns of spikes? If this is the case, what implications does this have for the view that the brain can be understood primarily in terms of spikes and spike-based plasticity rules?

16/09/08  11:00 - 12:00

ANC workshop talks: Edwin Bonilla and Mike Dewar, Chair: Steven Huang

Edwin Bonilla

Gaussian Processes for Multi-task Learning

Multi-task learning is an interesting scenario in machine learning where the learner aims to improve generalization by exploiting the shared information across different but related tasks. In this talk I will describe a class of multi-task learning models within the context of Gaussian processes. The main idea is that of, in addition to considering a shared covariance function on input features, task dependencies are modelled directly so that observations on one task affect the predictions on the others. This approach has been evaluated on different applications such as the compiler performance prediction problem and the exam score prediction task.

Mike Dewar

We can use spatiotemporal models to describe systems which exist across some physical space, and which change over time. In my talk I will focus on a first-order integro-difference equation based model. This is a popular form of model, due to its applicability to problems in ecology and geostatistics where it is used to model patterns that are difficult to represent with the more traditional reaction-diffusion equations.

I am interested in the model in an engineering context as it lends itself naturally to modelling systems using observations collected irregularly across space and regularly in time: a typical configuration when using sensor networks. My aim is to show that we can represent this spatiotemporal model using a standard linear dynamical system with a manageable state space. Once in this framework, we can use a standard EM algorithm to estimate the spatial field at each point in time, as well as the integral kernel which governs the dynamics of the system.

09/09/08  11:00 - 12:00

ANC workshop Talks: Judith Law and Roger Zhao, Chair: Adriana Teriakidis

Roger Zhao

Preliminary work on linking underlying mechanisms of face aftereffects with lower-level tilt aftereffects

Human perception of faces shows systematic aftereffects as a result of adaptation to specific faces [1], but it is not clear what mechanisms underlie these effects. We show that face aftereffects (FAE) can arise from Hebbian learning of connections in a LISSOM self-organising map model of visual cortex [2]. We suggest that, according to the modelling results, high-level aftereffects can be explained through the same mechanisms previously used for low-level effects like tilt aftereffects (TAE) [3]. We are working on linking this theory with experimental results of [1], and performing psychophysical experiment on male/female face trajectory to directly validate this theory.

[1] Leopold DA, OToole AJ, Vetter T and Blanz V. (2001). Prototype-referenced shape encoding revealed by high-level aftereffects. Nature Neuroscience, Vol 4 : pp8994.
[2] Miikkulainen R, Bednar JA, Choe Y and Sirosh J. (2005). Computational Maps in the Visual Cortex. New York: Springer.
[3] Bednar JA and Miikkulainen R. (2000). Tilt Aftereffects in a Self-Organizing Model of the Primary Visual Cortex. Neural Computation, 12(7):1721-1740.

Judith Law

"A cat and mouse game"

I am going to be talking about some work in progress that I am doing in collaboration with Tom Mrsic-Flogel at UCL. He is doing 2-photon imaging of mouse visual cortex and we are trying to identify the reason for the differences in organization of V1 in rodents and higher mammals. I have made a model of mouse V1 and cat V1 based on parameters from the literature and some of his new 2-photon data. The overall idea is that a similar overall cortical architecture in all species can lead to the different organizations because of differences in the parameterization.

26/08/08  11:00 - 12:00

ANC workshop Talks: Simeon Bamford and Jan Antolik, Chair: Graeme Phillipson

Simeon Bamford

Synaptic Rewiring for Topographic Map Formation

A model of topographic map development is presented which combines both weight plasticity and the formation and elimination of synapses as well as both activity-dependent and -independent processes. I statistically address the question of whether an activity-dependent process can refine a mapping created by an activity-independent process. A new method of evaluating the quality of topographic projections is presented which allows independent consideration of the development of a projection’s preferred locations and variance. Synapse formation and elimination embed in the network topology changes in the weight distributions of synapses due to the activity-dependent learning rule used (spike-timing-dependent plasticity). In this model, variance of a projection can be reduced by an activity dependent mechanism with or without spatially correlated inputs, but the accuracy of preferred locations will not necessarily improve when synapses are formed based on distributions with on-average perfect topography.

Jan Antolik

Neuronal selectivity and local map structure in visual cortex.
Nauhaus I, Benucci A, Carandini M, Ringach DL.
Neuron. 2008 Mar 13;57(5):627-8.

The organization of primary visual cortex (V1) into functional maps makes individual cells operate in a variety of contexts. For instance, some neurons lie in regions of fairly homogeneous orientation preference (iso-orientation domains), while others lie in regions with a variety of preferences (e.g., pinwheel centers). We asked whether this diversity in local map structure correlates with the degree of selectivity of spike responses. We used a combination of imaging and electrophysiology to reveal that neurons in regions of homogeneous orientation preference have much sharper tuning. Moreover, in both monkeys and cats, a common principle links the structure of the orientation map, on the spatial scale of dendritic integration, to the degree of selectivity of individual cells. We conclude that neural computation is not invariant across the cortical surface. This finding must factor into future theories of receptive field wiring and map development.


19/08/08  11:00 - 12:00

ANC workshop Talks: Chris Williams and Joanna Young, Chair: Tim O'Leary

Chris Williams

The PASCAL Visual Object Classes (VOC) Dataset and Challenge

I will discuss the lessons learned and issues arising from the The PASCAL Visual Object Classes (VOC) Dataset and Challenge.

Joint work with Mark Everingham, Luc Van Gool, John Winn, Andrew Zisserma

Joanna Young

Drosophila are capable of learning a number of associative tasks. A common method of testing learning and memory is using the T-maze. Flies are trained to with two different odours, one of which is reinforced with an electric shock, then they are moved to a choice point (testing) where they select between the same two non-reinforced odours. Time between training and testing can be varied depending on the phase of memory to be assessed. This method allows us to isolate memory mutants and to test for specific sets of neurons in the fly brain that are involved in memory acquisition, consolidation and retrieval.

In this talk I will give an overview of Drosophila learning and memory research to date and outline our project which aims to build upon these initial studies by assessing more complex learning tasks.

12/08/08  11:00 - 12:00

ANC workshop Talks: Matthias Hennig, Chair: Liana Romaniuk

Matthias Hennig

Modelling retinal waves - or what Minesweeper has in common with neural development

In this talk, I will give a summary of my recent work on retinal waves. Retinal waves are spontaneous, highly random, activity patterns in the developing retina (similar forms of activity are also observed in other immature brain structures). I have addressed the question which biophysical mechanisms are responsible for retinal waves with a computational model, and compared the simulations with data from multi-electrode recordings. It turns out that a very simple mean-field description can capture the behaviour of this model, which has some interesting analogies with the game of Minesweeper (thanks to Stuart Wilson for pointing this out!).

05/08/08  11:00 - 12:00

ANC workshop Talks: David Willshaw and Natasha Dare, Chair: John Davey

David Willshaw

Measuring maps of connections

I will review my ongoing work on how to measure the amount of order in a neural map. Such a map is defined by sampling at a given number of locations in one structure. For each case, the location in a second structure to which the sampled location projects is found.

Two situations are considered:

1. For maps which are completely ordered, if the sampling is done at a progressively finer and finer resolution, at some point the mapping will become disordered. Can the point at which the disorder arises be stimated from the ordered map?

2. For maps which are disordered, if some of the sampled locations are removed progressively from the mapping, at some point the map formed by the remaining set of sampled locations will become ordered. I give examples of this, for real maps.

I compare my method with other methods that have been proposed. I am hoping that this is the last 'presubmission' version of this talk.

Natasha Dare

There have been many models of isolated word recognition proposed, and as many models of eye-movement control during text reading, but few studies have explored how we recognise words when we are reading text. One way to start is with the effects that have been used to distinguish between models of isolated word recognition and investigate whether these effects also occur during text reading. One of these effects is transposed-letters priming which is the finding that there is more priming from an identical word whose letters have been transposed than one whose letters have been substituted: this rules out slot-based coding as an input system for word recognition models as this would predict equal priming from both. In this talk I propose a method for incorporating transposed-letters priming into an eye-tracking experiment and discuss how the results indicate that slot-based coding is not a viable input system for word recognition during text reading.

29/07/08  11:00 - 12:00

ANC workshop Talks: Graeme Philipson and Adrianna Teriakidis, Chair: Lawrence Murray

Adrianna Teriakidis

I will talk about the experiments I have been doing to determine whether, during the transition from poly-neuronal to mono-neuronal innervation in neonatal muscles, synapses are ever eliminated from muscle fibres in the absence of competition. This may seem like old news but I've got some new, possibly exciting, debatably accurate and definitely not statistically significant data which suggests that intrinsic withdrawal does occur. And I might add in a spinning picture too.

Graeme Phillipson

Binocular Depth Perception in a Human Subject with Bilateral Lesions of the Lateral Occipital Cortical Area

The human ability to discriminate binocular depth is most acute when humans are set the task of comparing the relative depth between two nearby visual features, with current evidence indicating that the ventral visual form pathway is responsible when the task requires a comparison between adjacent depth planes. We tested this hypothesis by measuring sensitivity to relative depth in the visual agnosic subject DF, whose major anatomical lesions are bilateral losses of the Lateral Occipital (LO) areas. Depth perception in DF was measured with a forced-choice, front-back task with a circular field of stereoscopic random dots as a function of their disparity. This central field was surrounded by a background field of random dots, which were always at zero disparity. Relative depth perception in normal subjects declines in acuity when an annular featureless gap is introduced between centre and background dots. The bigger the gap, the greater is the loss in acuity. However, DF’s acuity as a function of gap size showed essentially no change. DF’s stereoscopic depth vision is in many respects similar to normal subjects: she responds reliably to the cue of binocular disparity in random dot stereograms, is able to use binocular disparity to resolve ambiguous structure-from-motion images and reports a clear perceptual difference between correlated and anticorrelated binocular random dot figures. Results from DF are consistent with the proposal that relative depth between foreground and background visual features is processed through a pathway that passes through LO and forward to ventral visual areas.

The other people involved in this work were Jenny C.A. Read, A.David Milner, and Andrew J. Parker



22/07/08  11:00 - 12:00

ANC workshop Talks: Mark van Rossum and Wan-Yu Hung, Chair: Adam Chai

Mark van Rossum

What's next for computational neuroscience?

I will present some very preliminary ideas about a new effort to intergrate
computational neuroscience findings.

As we are thinking about writing a grant on this, feedback is very welcome

Wan-yu Hung

Learning plays as an important role in the development of synaesthetic associations

Synaesthesia is often described as an atypical communication between the senses (e.g., hearing a sound triggers a sense of colour; reading a word triggers a sense of taste). Studies have indicated that the development of synaesthetic associations may be traced back to early childhood experience. Once the synaesthetic links are established, they remain robust throughout the life and can provide as a frame for new associations to develop. Studies with multilingual synaesthetes have found that new synaesthetic associations for later-acquired languages tend to stem from the existing synaesthetic patterns in their first language. We tested a late Chinese-English bilingual by using both English letters and the Chinese bopomo (which is usually the first linguistic system taught in school before learning Chinese characters in Taiwan). In addition, we examined whether new synaesthetic associations can be acquired immediately. We recruited native English synaesthetes who are completely naïve to Chinese and tested whether they have any synaesthetic experience upon first seeing and hearing the pronunciation of bopomo symbols. In this talk, I will present our initial findings from two perspectives: (1) how bopomo symbols and English letters are coloured separately, and (2) whether the colouring of later-acquired language can be attributed to previous synaesthetic associations.

08/07/08  11:00 - 12:00

ANC workshop Talks: Adam Chai and Jesus Cortes, Chair: Lawrence York

Adam Chai

Multi-task Gaussian Process Learning of Robot Inverse Dynamics

The inverse dynamics problem for a robotic manipulator is to compute the torques needed at the joints to drive it along a given trajectory. We show that adaptive control handling different loads at its end effector naturally corresponds to a multi-task Gaussian process prior over the inverse dynamics. Experiments demonstrate that this multi-task formulation generally improves performance over either learning only on single tasks or pooling the data. This is joint work with Chris, Stefan and Sethu.

Jesus Cortes

Variations of coding accuracy by population of neurons after brief stimulus presentation

How the stimulus properties are represented by neural systems, ie. the neural coding, has been debated, and still today, for a long time. A striking feature difficulting those studies is that the neural coding seems to be adaptive, changing in a range of many different time scales. To address this problem, most of the electrophysiological research has been done at individual neurons. However, action, perception, learning and memory are believed to be encoded by population of neurons, and more specifically in different patterns of synaptic connectivity or neural wiring.

Very recently, the neural coding by population of neurons has started to be approached with experiments, by, for instance, using multi-electrode recordings or optimal imaging of population of neurons. However, a quantification of changes in population coding is difficult to address experimentally. It requires to collect enormous amount of data and this is an essential restriction to experimentalists. Alternatively, one can address this problem by modeling a population of neurons connected by a specific pattern of synaptic connectivity.

In general, after stimulus presentation individual neurons reduce their activity, they adapt and save metabolic expenses on coding a repetitive stimulus, while the population of neurons reproduce perceptual changes after adaptation, e.g. tilt after-effects.

I will present some results of how the population coding changes after brief presentation of general stimuli. Thanks to the modeling approach, I will discuss how the accuracy depends on different factors: 1) number of neurons in the population 2) different forms of adaptation, considering either synaptic or neural mechanisms 3) intrinsic variability of individual neurons.

Based on an analysis using Fisher information, stimulus discriminability increased for stimuli close to the adapting stimulus. These results suggest that visual adaptation is functionally advantageous from an information coding perspective and validate the "efficient neural coding hypothesis."

01/07/08  11:00 - 12:00

ANC workshop Talks: Lawrence Murray, Chair: Graeme Phillipson

Lawrence Murray

Modelling with continuous time stochastic processes

Many physical, biological and financial systems may be elegantly described using continuous time stochastic models. Such models are a mainstay of stock market prediction, neural modelling, and more recently in our own work studying the hemodynamics underpinning fMRI. Unfortunately, the continuous time setting poses additional challenges over that of discrete time, most significantly an intractable transition density in the general case.

In this talk I will highlight these challenges and discuss my own work in this area, particularly in the filtering, smoothing and parameter estimation problems for diffusion processes. I will introduce a proposed solution to all three problems, using a combination of stochastic Runge-Kutta and kernel density approximations.

10/06/08  11:00 - 12:00

ANC workshop Talks: David Sterratt and James Withers, Chair: Judith Law

David Sterrat

Making sense of censored data

In the developing visual system of many animals, maps of connections from the retina to targets such as the superior colliculus or optic tectum become more ordered over time. One way to measure this, as do our collaborators Ian Thompson and coworkers*, is to inject dye particles into a particular spot on the target. The dye is then transported retrogradely back down the axons to the cell bodies of neurons in the retina. The resulting pattern of dye spots in the retina shows which cells terminate in the target region in which the dye was injected.

In order to measure the degree of topography, we might want to plot the position of the projection onto the retina versus the position of dye injection in the target. However, there are at least two problems with doing this:
1. This pattern is often quite diffuse, and it is not obvious how to define the "centre" of the distribution.
2. There is missing or "censored" data. In order to save time when counting cell bodies, data is collected from one grid cell out of four.
Does this pattern of sampling affect the answer to (1) and can it be corrected for? I will describe my initial attempts to solve these problems. 


James Withers

Preserving thin structures in partial volume segmentation of brain MRI

The segmentation of MRI volumes of the brain into four constituent tissue types - gray matter, white matter, cerebrospinal fluid and blood/background - is necessary for tissue volume quantification, further analyses based on tissue type, and tissue-based constraints for other processing (such as white matter tractography). In this talk I will present my multi-scale approach for segmentation, and investigate the addition of deformable and pathological models. The ongoing development of new segmentation tools for labelling and visualisation will also be discussed. 

03/06/08  11:00 - 12:00

ANC workshop Talks: Irina Erchova and Michael Herrmann, Chair: Chris Palmer

Irina Erchova
ANC (School of Informatics) and Centre for Cognitive and Neural Systems (School of Biomedical Sciences)

Age related differences in the frequency-dependent properties of entorhinal cortex (EC) cells.

Cognitive ageing manifests itself as growing difficulty in performing some of the everyday behavioral tasks (though the deficits might be quite subtle). It reflects both loss of some essential cellular functions but also network adaptation to loss. Numerous biochemical and structural changes compromise neural functions and connectivity changes as a result of neural network adaptation to "new" activity patterns. One of the first brain areas to be affected by ageing process is the entorhinal cortex (EC), a cortical area feeding highly processed cortical input to hippocampus. As one of the current theories suggest that biochemical ageing is an extension of overall brain development, it is important to know how cellular properties develop with age. To this end, I compared electro-physiological properties of EC neurons in young (up to 5 weeks), adult (up to 12 months) and aged (24 months) rats, concentrating primarily on their frequency dependent properties, enabling behaviorally important theta rhythm generation in this area. The results indicate considerable age-related changes in the cellular properties of EC cells, with frequency preferences less prominent in both young and aged rats. I suggest a model linking these cellular changes to reorganization of the perforant pathway (PP), known to produce memory deficits.


Michael Herrmann

Self-organized criticality in neural networks with dynamical synapses

13/05/08  11:00 - 12:00

ANC workshop Talks: Peggy Series and Amos Storkey, Chair: James Withers

Peggy Series

Shannon versus Fisher Information in large populations of neurons.

Understanding how information is represented and transformed by populations of neurons is a major goal of neuroscience research. The accuracy with which information is encoded can be quantified using two types of measures: Shannon information (and derived quantities) and Fisher information (FI). Shannon information is commonly used to quantify the accuracy of single neurons or pairs of neurons. When populations of neurons are studied, however, FI is typically chosen, both in theory and in experiments.
I'll present some recent work, done with Ed Challis (MSc summer 07) looking at the relationship between these quantities in large populations of neurons.
More particularly, I am interested in the following question: what are the stimuli that are best encoded by a neuron? How does the 'information tuning curve' compare with the response tuning curve? How does it depend on the chosen information measure? The noise? The population size?

6th May 08

Bilal Malik

Using Drosophila as a model system to study the neural correlates of cognition has been proven to have many advantages over other model organisms. This is mainly due to its relatively simple nervous system and the many genetic tools available that enable us to cause genetic perturbation in a considerably manageable time. However, if Drosophila could be used in combination with higher model organisms like mice, it would provide an insight into the complexity of the molecular machinery that underlies cognition in humans. In this talk I will present our recent studies, carried out in collaboration with the The Wellcome Trust Sanger Institute Cambridge, on understanding the complexity of the synaptic proteome of these two model systems. I will also suggest that the evolution of this huge complex of proteins has taken a systematic course of evolution from yeast through to humans.

29th April 08

Chris Williams

Modeling image patches with a directed hierarchy of Markov Rrandom fields by S. Osindero and G. E. Hinton, NIPS*20 (Dec 2007)

BrainWave - a new approach to CNS drug discovery

Douglas Armstrong

Genome-wide human genetic studies aim to uncover genes linked to human psychiatric disorders. The earliest of these were funded by the pharmaceutical industry but these are largely being overshadowed by the academic sector. Preliminary results from multiple studies have uncovered few genes with large effects on disease but many with small, yet significant, effects. The drug discovery sector is currently looking at a huge problem, several hundred new genes (therefore potential drug targets) per disease. I will discuss how the CNS drug discovery pipelines work and how the BrainWave project aims to make an impact in this sector.

New data on binocularity and cortical magnification

Richard Shillcock

I will present some new data on binocularity in reading, together with new theoretical proposals for how the visual system is able to "zoom in" very literally on attended parts of the visual field.

Modelling Face Adaptation After effects

Roger Zhao

Human perception of faces shows systematic aftereffects as a result of adaptation to specific faces [1], but it is not clear what mechanisms underlie these effects. We show that face aftereffects can arise from Hebbian learning of connections in a LISSOM self-organising map model of visual cortex [2]. The model is trained and tested on faces from a generative model of a multi-dimensional face space [3]. Adaptation in the model shifts perception along a trajectory passing through the mean face. The perception of a target face on this trajectory is facilitated after adaptation to the target face's anti-face, but impaired after adaptation to other anti-faces.

In this talk I will start from the face space and generative model of a multi-dimensional face space. Then I will talk about Leopold's psychophysical experiment on face adaptation aftereffects. Last I'll show how I achieve face adaptation using a simple LISSOM model and face space generator, and how I measure aftereffects that qualitatively match Leopold's experimental results.

[1] Leopold et al. (2001) Nat. Neurosci 4:89-94. [2] Miikkulainen et al. (2005). Computational Maps in the Visual Cortex. Springer. [3] Hancock, P.J.B. Evolving faces from principal components. Behavior Research Methods, Instruments and Computers, 32-2, 327-333, 2000

On the design of a neuroinformatics project

John Davey

I am going to briefly outline my PhD work and a side project I have been working on. I plan to compare and contrast these projects, examine their experimental and computational components, and perhaps go as far as to draw some morals about the design of neuroinformatics projects in general.

For my PhD project, I have been attempting to identify RNA molecules in thalamocortical axons with a view to investigating microRNAs that may target these RNAs. I will talk about experimental work on identifying RNAs and computational work on predicting microRNA and RNA targets.

I will also describe some computational work on modelling the growth of dendrites in barrel cortex.

15th April 08

Andrew Dai

The problem of learning aspects and relations of a rich but noisy dataset continues to be a sought after goal as machine learning models improve in their ability to extract rich structure from low-level features of the data. Much of current literature assumes that data is represented by points in some high-dimensional space, thus allowing standard tools to be used on the vector representation of the data. However, this hides the structure of the underlying data and chooses to ignore the structure by assuming the attributes of the data are IID. This limits the models to learning flat propositional rules that are limited in scope as opposed to those rules and worlds that the richer models of first order logic are able to capture. The aim of this project is to address the possibility of discovering multiple levels of relational structure from data with developing hierarchical models and algorithms to perform tractable inference and learning on them.

8th April 08

Jim Bednar

8th April 08

Jan Antolik

1st April 08

David Willshaw

1st April 08

Steven Huang

Topographic and laminar models for the development and organisation of spatial frequency and orientation in V1

Chris Palmer

Primary visual cortex (V1) in higher mammals contains smooth topographic maps for visual features like orientation (OR). These maps typically consist of vertical columns of neurons with similar feature preferences. Neurons have also been found to have different spatial frequency (SF) preferences, and a consensus is emerging that in the superficial layers SF is mapped continuously across the cortical surface. However, other evidence suggests that SF may differ systematically with cortical depth, at least in layer 4, where the M and P pathway afferents terminate at different sublaminae. It is not yet clear whether the topographic organization for SF observed in the superficial layers is maintained throughout the input layers as well, or whether there is a switch from a laminar to a topographic organization along the vertical dimension in V1.

I present results from two alternative models of how V1 organizes in response to input from multiple spatial frequency channels in the LGN, differing on whether they develop laminar or topographic organisation in layer 4. Both models lead to topographic organization at higher layers, and are consistent with current experimental evidence, but they make very different predictions that can be tested in future experiments. The results suggest that Hebbian learning mechanisms can explain the development of laminar and topographic organisation for spatial frequency, as long as there are mechanisms for explicitly pruning or selecting between different SF channels during early development.

25/03/08 Topographic and laminar models for the development and organisation of spatial frequency and orientation in V1

Adrianna Terikidis

Given that the initial innervation of the neuromuscular junction is random, it would be expected that sibling axons (from the same motor neuron) converge on the same endplate. My results show that this does occur and that they seem to be competitively eliminated. This raises the issue of how they compete with each other since known mechanisms of competition rely on inter-neuronal differences. Also does this mean that initial innervation is indeed random? I will go through my results and present some work by other people relating to the questions raised.

18th March 08

Judith Law

Throughout the development of the map of orientation preference in primary visual cortex (V1), the statistics and strength of afferent inputs, synaptic strengths and sizes, and lateral connection patterns are all changing. It is therefore surprising that despite these massive changes, the orientation map develops with a remarkable stability in both orientation preference and orientation domain size (Chapman et al., 1996).

It has been proposed that homeostatic mechanisms that automatically adjust the intrinsic excitability of neurons and/or multiplicatively scale synaptic strengths allow individual neurons to maintain stability of overall activity levels (Desai, 2003, Turrigiano and Nelson, 2004). We have explored the possibility that homeostatic plasticity also underlies the stability of map organization.

Several homeostatic rules have been used to replace ad-hoc methods in simple self organizing map (SOM) models (e.g. Butko and Triesch, 2007, Sullivan and de Sa, 2006). However it still remains to be seen whether these rules will be sufficient in a more realistic model that includes adapting lateral connections.

I will discuss a more realistic model (based on the LISSOM architecture) which can reproduce stable orientation map development by maintaining stable neuron responses in V1 while retaining the balance between input types. Homeostatic regulation of intrinsic excitability and synaptic weight normalization can maintain stable average activity levels, however it is also necessary to maintain a balance between the strengths of afferent and lateral (excitatory and inhibitory) inputs to V1.

18/03/08 Multi-scale segmentation of dual-channel MRI using super-resolution and tubular structure detection

James Withers

Brain tissue segmentation is complicated by noise and partial volume voxels that contain mixtures of two or more tissue types. Here, I present three extensions to segmentation methods which may allow more accurate classification of small-scale structures while still providing robustness to noise. Firstly, the location and orientation of fine tubular structures can be identified using differential geometry. Subsequent neighbourhood filtering operations, for ensuring local homogeneity or performing volume super resolution, may then operate with different scales and orientations according to the structure of interest. Finally, dual-channel T2-PD segmentation can be improved by weighting the information from each modality to better emphasise contrasting tissues. Quantitative results are shown using simulated volumes and qualitative assessments from a real dataset are presented. Finally, some future applications to brain extraction, image registration and an Expectation-Maximisation segmentation framework are discussed.

05/02/08 Shannon Information Capacity of Discrete Synapses

Adam Barrett

There is some evidence to suggest that a realistic model of synaptic plasticity and memory storage should involve synapses that have only a fixed number of discrete weight states. Networks equipped with such synapses behave very differently from networks with synapses whose weights are unbounded and may vary in a continuous manner. In particular, the over-writing of old memories by new memories is a feature of such models. I present a framework for studying the storage capacity of discrete synapses in terms of the Shannon information capacity of a single integrator neuron. For optimal learning rules, I discuss the rich dependence of information storage on the number of weight states, the number of synapses, and the sparseness of input patterns presented. I briefly contrast these findings with results from other models.

29/01/08 Synaesthesia in Chinese: A perspective from Chinese characters and their associated phonetic spellings.

Wan-Yu Hung

Synaesthesia is an unusual phenomenon in which different senses are combined in an atypical way; for instance, the sounds of music induce colour vision; words trigger tastes, etc. This talk focuses on synaesthesias that are related to languages, particularly the so-called ?word-colour synaesthesia? that describes an instinctive colour experience triggered by words. One major finding of recent studies in English has pointed out that for English word-colour synaesthesia, a critical trigger might consist in the initial letter or the initial vowel of the word. It is found that this particular pattern resembles much the processing of word recognition. Thus, we speculate that this type of synaesthesia might be rooted in one?s language environment and thus should be language specific. In this talk, I will discuss this from the perspectives of Chinese synaesthesia based on our findings with 6 Chinese speaking synaesthetes (native Chinese: n = 3; non-native Chinese: n = 3). This involves examining whether their colours would be related to (a) the initial letters, (b) the initial vowels, or (c) the lexical tones of characters. Additionally, since each character has its own associated phonetic spellings, which are usually taught before learning character orthography, we ask whether the above linguistic effects would also be present in the explicit presentation of the spellings. I will discuss these findings in relationship to native vs. non-native speakers and Chinese characters vs. their associated spellings.

29/01/08 Population code and rapid adaptation

Jesus Cortes

The cellular and synaptic mechanisms underlying visual adaptation have been debated during the last years and even today. Contrarly to spike-frequency adaptation, synaptic mechanisms are responsible of rapid adaptation, occurring at the time scale of the stimulus presentation (typically, various hundreds of milliseconds). Based on these results, we present here a mechanistic model for rapid visual adaptation. The model considers a stantard recurrent model of V1 (ring model) in presence of short-time synaptic depression affecting both excitatory and inhibitory intra-cortical synapses. The computational model reproduces post-stimulus changes at the single-neuron and population levels. After adaptation individual neurons reduce their activity, thus saving metabolic expenses on coding a repetitive stimulus. Variations on the population of neurons reproduce perceptual changes after adaptation, e.g. tilt after-effects.

A clear advantage of a mechanistic model is that one can quantify how rapid adaptation affects the neural coding to both individual neurons and population. Such precise connection is difficult to establish, because noise correlations affects the specific encoding of sensory stimuli. Furthermore, how adaptation changes the structure of correlations among individual neurons is difficult to address experimentally. This is the significance of using a computational model to explore how adaptation changes the population coding, responsible of changes at the perceptual level. By using information theory, we validate the ?efficient neural coding hypothesis?: after adaptation noise correlations decreases and stimuli discriminability increases at the population level i.e. Fisher Information increases for a given stimulus close to the adaptor. These results suggest that visual adaptation is functionally an advantageous, from an informational perspective.

(Other collaborators. Daniele Marinazzo, Peggy Series, Terry Sejnowski and Mark van Rossum)

22/01/08 Multiple time scales in synaptic transmission: nonlinearities and speculations

Matthias Hennig

Synaptic transmission is not a static process, but known to have dynamic components, which change the synaptic efficacy in a stimulus-dependent manner. This is known as synaptic short term plasticity (STP). It is further known that synapses have a variety of different regulatory processes acting at different sites and on many different temporal scales. Current models of STP however typically assume contributions from just one or two processes, hence only give a strongly simplified picture of synaptic transmission. I will discuss a model of a synapse that is more complete and complex in this respect, and show that it yields a more complete description of various experimental findings. I will then discuss nonlinear properties of this model, which are the result of simultaneous dynamics on multiple time scales, and speculate about some functional implications.

22/01/08 Coupled Markov Models to Learn Handwriting Primitives

Ben Williams

I'll talk about my phd project, firstly an overview of the model, then a walk-through of the generative procedure, then I'll show how handwriting style can be captured by the primitives in our model, and then finally a comparison with a similar model.

Here is a more formal abstract: Biological movement is built up of sub-blocks or motor primitives. Corresponding data such as handwriting must also therefore be decomposable into primitives. Inference of the shape and the timing of primitives can be done using a Factorial HMM based model, allowing the handwriting to be represented in primitive timing space. This representation provides a distribution of spikes corresponding to the primitive activations, which can also be modelled using HMM architectures. We show how the coupling of the low level primitive model, and the higher level timing model during inference can produce good reconstructions of handwriting, with shared primitives for all characters modelled. This coupled model also captures the variance profile of the dataset, which is accounted for by spike timing jitter. With a shared set of primitives, the spike timing information forms a compact code for the character class, as well as the precise variation of the reproduction, in the nature of handwriting. Generating a timing code without an explicit timing model produces a scribbling style of output.

15/01/08 EM and some variants

Kian Ming

Although the Expectation-Maximisation algorithm is a convenient way to find maximum likelihood estimates of parameters, it is known to converge slowly. In this talk, I will discuss some existing variants of the EM algorithm --- PX-EM, Overrelaxed EM and triple jump EM --- that I have found to be effective in increasing the convergence speed on the multitask Gaussian process regression model. I will also highlight the effectiveness of staged-optimization in parameter estimation.

8th January 08

Tim O'Leary

In this talk I shall outline some experimental work on the effect chronic depolarisation has on immature central neurons. Specifically, the work indicates that neurons posses the ability to regulate their 'intrinsic excitability' or 'propensity to fire an action potential' in response to variations in their resting membrane potential. The key result is that the voltage-gated sodium conductance is homeostatically regulated, but the voltage-gated potassium conductance is not.

A simple computational model has proven to be effective in planning experiments and tackling hypothetical 'what if?' questions surrounding the work. In particular, the model I shall describe indicated that voltage-gated potassium currents were indeed unlikely to play a decisive role in modulating excitability. On the downside, the task of fitting the model was quite painful, and the number of parameters bestowed with 'ballpark' or 'reasonable' values suggests that it has predictive power on a somewhat shallow level.

As an introduction, I shall present what I believe to be several very interesting facts about the role of ion transport in single-celled organisms and argue that evolution offers a clue as to why voltage-gated potassium channels don't seem to play a part in modulating excitability.

04/11/07 Uncertainty drives decision making via overlapping cortico-subcortical channels: Imaging and effective connectivity supports and informs computational models.

Liana Romaniuk

The efficient selection of actions appropriate to a given context depends on the effective suppression of competing alternatives. In the absence of a simple relationship between context and action, multiple action strategies must be evaluated by the brain, and a particular behaviour executed at the expense of other possibilities. The world available to us perceptually can never be thought of as complete, forcing us to deal with the ambiguities arising from missing information. Ambiguity is known to engage medial, dorsolateral, orbital and ventrolateral frontal cortex, inferior parietal areas, insula, amygdala, striatum and thalamus, though it is unclear how this distributed network acts in concert to direct useful motor output. Proponents of centralization favour the basal ganglia (BG) as being optimally configured to arbitrate accordingly. The BG act as a key way station along corticothalamic loops, exerting a suppressive influence via partially segregated, closed loop "channels". Given sufficient cortical input, this is said to become selectively disinhibitory, thus releasing the ?brakes?and permitting action execution. While attractive, this has yet to be demonstrated in vivo in humans. It is safe to say that uncertainty will evoke interference between competing responses, making it more essential to suppress the losing alternative. That BG activity is modulated by uncertainty is not in any doubt, but it is our aim here to tease apart the closely related functions of uncertainty and suppression within the BG output nucleus ? globus pallidus (GP). Using a novel behavioural paradigm, fMRI, bivariate Granger causality and computational modelling, we demonstrate that the BG integrate multiple decision variables over distinct channels, and determine its relationship with expressed behavior.

04/11/07 Computational Neuroanatomy of Drosophila melanogaster

Mark Longair

In this talk I will describe some of our attempts to test standard neuranatomical hypotheses about the structure of the fruit fly brain using image analysis tools. This includes:
- Using our collection of image stacks of enhancer trap lines to examine the layered structure of the fan-shaped body.
- Some updates to my previous talk on semi-automated tracing of neurons in 3D image stacks.
- Adding polarity information to our connectivity data using a reporter construct that is preferentially distributed to dendrites.
In addition I will discuss some different image registration methods that I have been using and a new tool for fine-tuning the placement of landmark points in an image stack.

20/11/07 The Central Complex (CC) of Drosophila

Jo Young

The Central Complex (CC) of Drosophila has been implicated in behavioural activity and is thought to act as a processing centre in the brain. To date, several studies have been conducted on the behaviour of the CC however few have addressed the development and connectivity of this structure. In order to understand the functions of the CC it is critical to determine how it is connected. The CC is an exclusively imaginal structure that develops during early metamorphosis. We show that the study of connectivity can be facilitated through the analysis of the developing fly brain and by analysis of Central Complex mutants.

We have isolated a set of P{Gal4} enhancer trap lines expressed in the Central Complex, focusing on the Fan Shaped Body. We have then used these lines to trace the development of specific neurons every four hours throughout metamorphosis and have subsequently generated several complete sets of developmental series data. From this we have deduced a timeline of development for specific neurons of the Central Complex.

Finally, we have used these P{Gal4} lines to analyse a set of known Central Complex mutants. Drosophila Central Complex mutants have been used frequently in the past for behavioural studies and have several structural abnormalities in addition to locomotor deficits. No detailed analysis on these mutants has been performed. Using the P{Gal4} enhancer trap lines we can analyse the disorganised neuropil of the CC mutants in detail. This information could be used in future behavioural studies to potentially reveal functions of individual sets of neurons.

20/11/07 A brief tutorial on bifurcation analysis as applied to reduced neuron models

David Sterratt

The Hodgkin-Huxley model of squid axon contains four coupled differential equations in each compartment, and when more types of ion channel are modelled, the number of differential equations in a compartment increases. Understanding the possible types of dynamics exhibited by this type of model (e.g. burst firing, tonic firing, chaotic firing) is hard.

Reduced models of neurons such as the Fitzhugh-Nagumo model or Morris-Lecar model usually contain two or three differential equations, and yet retain many of the dynamical features of the more complex models. They are also amenable to phase plane analysis and bifurcation analysis.

Hitherto one prevalent computer tool for bifurcation analysis has been XPPauto, which although powerful, has some drawbacks. Recently I have been using a Python-based tool, PyDSTool, which has the advantage of being programmable. I will give a brief demonstration of using one of the tools to do the bifurcation analysis of a reduced neuron model.

13/11/07 Somatosensory perception and illusions

Irina Erchova

This is an ?extra? presentation that is not directly related to my current work, but is vaguely related to my previous research on neural correlates of somatosensory perception. I have been recently researching the topic for upcoming neuroscience public event. I will introduce a number of well-known somatosensory illusions and their current interpretation in terms of evoked neural activity and a Bayesian perceptual model. I will also talk about cross-model visual-somatosensory interactions and illusions resulting from the conflict information delivered by different systems.

13/11/07 Evolution and anatomical specialisation of synapse proteome complexity

Andrew Pocklington

06/11/07 Does the brain know when it changes?

Peggy Series

The properties of sensory neurons are not fixed. They change dynamically according to the spatial and temporal context and the task being performed. Adaptation, for example, is known to result in a decrease in response amplitude, while attention enhances responsivity. How does the rest of the brain interpret these changes? Does the read-out adapt at the same time as the sensory neurons?

We explore this question in the context of sensory adaptation, focusing on the examples of motion direction adaptation and contrast adaptation. In our framework, perception is modeled as resulting from an encoder-decoder cascade. The encoder corresponds to the response properties of a population of cortical neurons and changes during adaptation. Different types of decoders are considered, which are either fixed and 'unaware' of the adaptation state, or which change dynamically at the same time as the encoder, being thus always optimal and 'aware' of the adaptation state. Their predictions are compared with the psychophysical data for estimation and discrimination tasks.

We find that simple models of neural adaptation coupled with 'unaware' read-outs can account for the main features of the psychophysical results. We discuss the significance of having 'unaware' read-outs, and their relevance in other phenomena, such as contextual interactions, attention and perceptual learning.

Work in collaboration with E. Simoncelli and A. Stocker (NYU).

06/11/07 Multi-task Gaussian Process Prediction: (Joint work with Kian Ming A. Chai and Chris Williams)

Edwin Bonilla

In this work we investigate multi-task learning in the context of Gaussian Processes (GP). We propose a model that learns a shared covariance function on input-dependent features and a ``free-form'' covariance matrix over tasks. This allows for good flexibility when modelling inter-task dependencies while avoiding the need for large amounts of data for training. We show that under the assumption of noise-free observations and a block design, predictions for a given task only depend on its target values and therefore a cancellation of inter-task transfer occurs. We evaluate the benefits of our model on two practical applications: a compiler performance prediction problem and an exam score prediction task. Additionally, we make use of GP approximations and properties of our model in order to provide scalability to large data sets.

30/10/07 Loading on Phonology

Marina Papoutsi

In this sort talk, I will present some results from an fMRI study looking at some of the details of phonological processing in the brain. We used tightly controlled phonological words, manipulating only speficic dimensions of the stimuli, to map the brain centers responsible for the processing of these particular features. The results will be discussed in the context of what we already know about phonological processing.

16/10/07 Predicting spiking rates of neurons in inferotemporal cortex

Chris Williams

I will describe work carried out by Fabian Wauthier this summer under my supervision, with the aim of predicting the responses of neurons in inferotemporal cortex to presented images. The work is carried out in conjuction with Peter Foldiak (St Andrews).

16/10/07 Continuous Time Particle Filtering for fMRI

Lawrence Murray

We construct a biologically motivated stochastic differential model of the neural and hemodynamic activity underlying the observed Blood Oxygen Level Dependent (BOLD) signal in Functional Magnetic Resonance Imaging (fMRI). The model poses a difficult parameter estimation problem, both theoretically due to the nonlinearity and divergence of the differential system, and computationally due to its time and space complexity. We adapt a particle filter and smoother to the task, and discuss some of the practical approaches used to tackle the difficulties, including use of sparse matrices and parallelisation. Results demonstrate the tractability of the approach in its application to an effective connectivity study.

02/09/07 Improving partial volume segmentation in MRI of the brain

James Withers

The segmentation of images of brain tissue acquired by magnetic resonance imaging (MRI) is complicated by partial volume (PV) voxels that contain a mixture of two or more tissue types, as well as by noise. I will examine three techniques for PV segmentation which extend existing methods in order to more accurately quantify voxel mixtures whilst providing robustness to noise. Taking a parametric statistical image model where voxels are formed from contributions of one or two tissue types, classification is subsequently performed on a super-resolved image. An expectation-maximization approach is used to simultaneously estimate the parameters of the model based on selected sets of voxels and perform a PV classification, which is influenced by a priori knowledge of vessel-like structures in the image. Quantitative results on simulated images and qualitative assessments by experts on real MR images of the brain are presented, alongside comparisons with state-of-the-art tools.

02/09/07 Parafoveal-on-foveal effects: a test of serial v. parallel processing during text reading

Natasha Dare

Models of text reading seek to reproduce the well-established patterns of fixations, saccades, regressions and word skipping that occur when the eyes take in written information. A current controversy that divides modellers is whether these models should rely on serial word processing, in which each word is processed in turn, or allow for parallel processing, in which more than one word can be processed at a time. There are many consequences of these different approaches, including whether or not they predict the presence of parafoveal-on-foveal effects when the properties of upcoming words affect processing of the current word. I present details of my latest experiment which seeks to establish a novel parafoveal-on-foveal effect of orthographic priming from an upcoming word and thus provide further evidence for the appropriateness of the parallel approach.

04/09/07 Rhabarber - a new programming language

Stefan Harmeling

Mikio Braun and Stefan Harmeling are working on a new programming language that should be useful especially for machine learning researchers. While being designed to look like the C programming language, Rhabarber is as feature-rich as LISP. This talk will give a live demo of some of its features and discuss future plans.

04/09/07 Topographic and laminar organisation of spatial frequency and orientation in a developmental model of V1 using natural images

Chris Palmer

Spatial frequency is an essential component within the composition of the visual image, and as such must play an important role in building a picture of the world which visual organisms use to interpret their environment.

It is well known through electrode penetrations studies that cells in the visual cortex can have preference for different spatial frequencies, but it is not known if or how this preference is organised within cortex. A number of experimental studies have tried to address this question, but they are at odds with one another, and the answer remains uncertain.

By examining this question using a Hebbian based computational model with a biologically plausible structure, we are able to examine spatial frequency (SF) organisation, and have demonstrated the plausibility of SF preference having a continuous laminar organisation in layer 4C of V1. This is a consequence of representing the output from lateral geniculate nucleus (LGN) as two pathways, the M and P pathway (with a lower and higher SF preference, respectively), these pathways project distinctly to different laminar layers within V1. This results in realistic matching orientation maps, with differing SF preferences. These models have been trained successfully using gaussian input patterns and natural images.

The range of SF that can be represented in cortex is restricted in the model by the range of SF preference within the LGN. By extending this range within the model, we are able to reproduce the range of preference in cortex that is found experimentally. We also find that cortical units must be constrained to receive input from only LGN cells of similar size, otherwise the range of SF preference found in V1 is noticeably narrowed.

28/08/07 Orthographic satiation in Chinese

Richard Shillcock

I will report a provisional analysis of data from Nien-Chen Lee, an MSc student. Two experiments were carried out on the little-researched phenomenon of "orthographic satiation" in Chinese. When native speakers of Chinese stare at a single character for half a minute or so, they frequently experience a satiation effect, in which there is a loss of gestalt of the figure - it becomes odd, disorganized and loses something of its meaning. We carried out two experiments that replicated this effect and showed significant effects of the sex of the reader, the functional cortical lateralization of the reader, and the left-right structure of the character. These effects provide insights into the nature of Chinese character recognition and have broader implications for experiments on visual word recognition.

28/08/07 Keeping Hebb under control - homeostatic regulation of intrinsic excitability in hippocampal neurons

Tim O'Leary

The favourite candidate mechanism for learning and memory in the mammalian CNS is Hebbian learning - a process which tends to strengthen synapses between excitatory cells in response to correlated firing patterns. While there is abundant experimental evidence in support of a synaptic Hebbian paradigm, there is at least one other non-synaptic mechanism for implementing Hebbian learning: modulation of intrinsic excitability. Numerous in-vitro and in-vivo studies have shown that persistent changes in an individual cell's response to depolarising current can be induced by elevated input, whether imposed synaptically or by direct injection of current (Moyer Jr et al., 1996; Cudmore and Turrigiano, 2004; but see Daoudal and Debanne, 2003 for a review).

Both forms of Hebbian learning are unstable in isolation. Synaptic learning rules require additional mechanisms such as 'synaptic scaling' to prevent populations of strong synapses entering a positive-feedback loop and weak synapses diminishing completely. Historically, these mechanisms were introduced ad hoc into artificial neural networks before experimental evidence alluded to their existence (Miller and MacKay, 1994; Abbott and Nelson, 2000).

Likewise, a Hebbian-type learning rule for modulating intrinsic excitability is unstable without some kind of homeostatic regulation. In this talk I will describe a novel in-vitro model for homeostatic control of intrinsic excitability. We find that cultured hippocampal neurons respond to chronic depolarisation over a period of days by attenuating their response to injected current. Cells subjected to treatment exhibited a marked increase in the amount of steady current required to induce spiking (by at least an order of magnitude). This effect was found to depend on the level of depolarisation and the length of treatment, and appears to be accompanied by a change in voltage-gated potassium currents in the cells. Consistent with these observations is a prominent hyperpolarising shift in resting membrane potential (of more than 10mV) and a decrease in input resistance (to approximately 25% of control values).


Abbott LF, Nelson SB (2000) Synaptic plasticity: taming the beast. Nat Neurosci 3 Suppl:1178-1183. Cudmore RH, Turrigiano GG (2004) Long-term potentiation of intrinsic excitability in LV visual cortical neurons. J Neurophysiol 92:341-348. Daoudal G, Debanne D (2003) Long-term plasticity of intrinsic excitability: learning rules and mechanisms. Learn Mem 10:456-465. Miller KD, MacKay DJC (1994) The role of constraints in Hebbian learning. Neural Computation 6:100-126. Moyer Jr JR, Thompson LT, Disterhoft JF (1996) Trace Eyeblink Conditioning Increases CA1 Excitability in a Transient and Learning-Specific Manner. Journal of Neuroscience 16:5536.

28/08/07 Homeostatic plasticity in a model of V1 orientation map development"

Judith Law

Homeostatic plasticity is the proposed regulatory system by which individual neurons maintain stability of their overall activity levels despite changes due to Hebbian plasticity at specific synapses (reviewed in Desai, J Physiol Paris 2003 97: 391-402). It is considered to be an important complement of Hebbian plasticity, especially during learning and development in the nervous system.

During development in mammalian cortex, neurons become selective for various features of the visual input, typically forming smooth maps of feature preference across the cortical surface e.g. the map of orientation preference in primary visual cortex (V1). We have explored several proposed homeostatic plasticity rules within a developmental topographic map model based on the LISSOM architecture (Mikkulainen et al., Computational Maps in the Visual Cortex, Springer 2005). In the context of learning in a two-dimensional map, the statistical patterns of activation can be quite different from those of a single neuron simulated in isolation. It is important to know whether homeostatic mechanisms proposed for single neurons are sufficient to cope with this greater variability, which provides a constraint on the types of mechanisms that could be effective in real systems.

Using the rule proposed by Triesch (Neural Comput. 2007 19: 885-909) to create a less complex but more robust version of LISSOM, we show that the model can develop orientation maps that maintain a consistent structure throughout development, as found in chronic optical imaging of ferrets (Chapman et al., J. Neurosci 1996 16: 6443-6453).

I plan to discuss these results and highlight some of the outstanding issues with this work. I shall also present a possible new homeostatic rule which allows for different levels of homeostatic plasticity throughout the cortex, depending on the previous history of activation. In this new implementation, the neuron responds to changes in excitability much more quickly and is thus more effective in regulating the neuronal firing rate.

21/08/07 A state based model of Long Term Potentiation and Depression

Guy Billings

Long Term Potentiation and Depression (LTP/D) is a robustly observed neural phenomenon. LTP/D allows the synaptic weight connecting neurons to be changed. Therefore LTP/D is thought to be the biological substrate of memory and the biological basis of the 'learning rule'. Intriguingly synaptic changes due to LTP/D appear to be able to persist on multiple timescales. This raises a question: Why have multiple timescales of synaptic plasticity?

In this talk I will present a state based model of LTP/D that I have developed in collaboration with Mark van Rossum and Adam Barrett. Our model of LTP/D explicitly takes into account the multiple timescale feature of LTP/D. I wish to use this model to demonstrate that a consequence of the multiple timescales of LTP/D, is that synapses in the hippocampus could be 'overloaded'. This overloading might allow us to store information on short timescales with only temporary disruption to patterns stored on long timescales. This feature might a be a method of accommodating both long term storage and rapid learning within a single population of synapses.

21/08/07 Modeling the development of tri-chromatic color maps in V1

Jim Bednar

(Joint work with Judah B. De Paula and Risto Miikkulainen of the University of Texas at Austin)

Neurons in the primary visual cortex of primates form topographic maps organized around visual features such as orientation, direction, and eye of origin. Experiments suggest that the cells selective for color are organized into small, spatially separated blobs, in striking contrast to the large, spatially contiguous patterns typical of other preference maps. We have constructed a self-organizing tri-chromatic model of V1 that helps explain why color maps are so different from others. Neurons in the model are initially unselective, and develop multi-lobed ON/OFF receptive fields through Hebbian learning of color images of natural scenes processed by photoreceptors and retinal ganglia. The model develops realistic color-selective receptive fields, color maps, ocular dominance maps, and orientation maps. Color-selective blobs are located inside ocular dominance columns, and lateral connections link cells with similar orientation preferences, matching previous experimental results. Further, the model makes a number of predictions that can be tested in future experiments:

1. The color map has three types of color-selective blobs and a unique cortical activation pattern exists for each of the pure color hues.

2. The empirical blob-like organization for color emerges as long as (a) the training images have a higher brightness contour gradient than hue contour gradient, and (b) the inputs are highly correlated between the eyes. Otherwise, the color blobs regularly extend across borders of ocular dominance stripes (contrary to macaque results).

3. Neurons in areas where red and green patches are near each other respond to both red and green, causing them to maximally prefer yellow, even though there are no yellow photoreceptors in the model (or animal) retina.

4. Long-range lateral connections between color-selective cells are specific to color: Blue-selective neurons connect to blue selective neurons, red-selective to other red-selective neurons, and so forth.

Thus the model replicates the known data on the organization of color selectivity in macaque V1, gives a detailed explanation for how this structure develops and functions, and provides concrete predictions that can be tested in future experiments. These findings suggest that a single self-organizing system may underlie the development of orientation selectivity, eye preference, color selectivity, and lateral connectivity in the primary visual cortex.

14/08/07 Modelling spontaneous activity in the developing retina

Matthias Hennig

Patterned spontaneous activity is a common feature of developing neural systems. In the immature retina, spontaneous activity has the form of propagating waves, and it has been demonstrated that it is required for the correct development of the retina and the structures in the higher visual system. In this talk, I will present a model for retinal waves and show how different activity-dependent processes can interact to increase the variability of the observed activity patterns.

14th August 07

Steve Huang

The head direction cells of the rats are only active when the rats face specific directions of the environment and can maintain their directional specificity with only vestibular input. They can thus provide directional information to the rats when the external cues are absent. They also reference their receptive fields with respect to prominent cues of the environment, and their preferred orientation in the environment can be changed by rotating these cues. However, sometimes the change of the environment can give rise to conflicting directional information between external cues (e.g. when one cue rotates by 90 degrees clockwise and another by 90 degrees anti-clockwise). Under such situation, the head direction cell network is forced to resolve the conflict as the experimental evidences all suggest that these cells always behave coherently together. In my talk I will present the data from simulations that was designed to explore the behaviours of the network under this kind of conflict and also experimental data from the recordings of the place cells when the rats were given the conflicting cue information.

07/08/07 Subthreshold resonance underlies history dependent excitability: a novel biophysical mechanism for working memory.

Fabiano Baroni

How single neurons encode information is a fundamental question in neuroscience. In particular, is the overall firing rate what carries the information or is the timing of single spikes also important? Several experimental studies have found spatiotemporal sequences of spikes which occur with millisecond precision. The significance and hence the informational capabilities of these precise temporal sequences can be assessed by statistical methods, but which are the candidate dynamical mechanisms at the single neuron level which could read a precise temporal code? Traditionally the detection capabilities of neurons have been assessed by observing how their output changes as a function of their input. Hence two stimuli are considered to be indistinguishable as long as they both result in a single postsynaptic spike, or in the lack of it. We show that this is not necessarily the case: two different input stimuli can still be distinguished if they bring the neuron to a different internal state, and hence change its response properties to future stimuli. By defining neuronal excitability as the minimum strength of a suprathreshold input, we lump all the intrinsic variables of a neuron model in a single, physiologically relevant quantity. This framework allows us to quantitatively assess the temporal detection capabilities of different neuron models. We found that subthreshold resonance on multiple time scales allows the discrimination of temporal intervals which are far back in the input history, and hence implements a dynamic encoding mechanism which might be important in network behavior. This mechanism can coexist with short term synaptic plasticity as a substrate for working memory.

31/07/07 Storage with discrete synapses

Mark van Rossum

There is some evidence that synapses have only a certain number of states. This has sparked theoretical interest how this affects memory storage. In contrast to earlier Signal-to-noise analysis [Fusi Abbot 07], we analyze this question using information theory and find some interesting results.

31/07/07 Learning topological relationships for white matter pathways

Jon Claydon

The major white matter fibre tracts are present in all healthy brains, but identifying a particular tract in magnetic resonance imaging (MRI) data obtained from groups of subjects is challenging. The development of algorithms for fibre tracking using diffusion MRI has helped, but reconstructed tracts are sensitive to the algorithm's initialisation and affected by noise in the data. This work describes attempts to model the expected topological relationships between comparable tracts, and then to use the model and a reference tract to improve segmentation consistency.

24/07/07 Combining gene expression data sets from different experimental conditions.

Adriano Werhli

The assumption so far has been that the molecular biological system of interest can be characterized by a unique regulatory network. What we are actually aiming to infer, though, are the active parts of this network, which may differ under different experimental conditions. To illustrate this point, consider a transcription factor that potentially upregulates a group of genes further downstream in the regulatory chain. If the experimental conditions are chosen such that the gene coding for this transcription factor is never expressed itself, then the respective subnetwork will never be activated, and hence cannot be inferred from the data.

When aiming to reconstruct a network from gene expression profiles taken under different experimental conditions, there seem to be two principled approaches we may pursue. The first is to ignore the changes in the experimental conditions altogether and merge the data into one monolithic set. The problem with this approach is that it inevitably blurs the differences between the different conditions and thereby obscures the biological insight we are aiming to gain. The second approach is to keep the data obtained under different conditions separate, and to infer separate regulatory networks active under these different conditions. Breaking a sparse data set up into smaller units will inevitably increase the uncertainty about inferred network structures.

In the present work we aim to pursue a compromise between the two extreme procedures described above. We first introduce the new approach and then show the results we obtain when applying the new method to simulated and real data. We finish with a discussion about the results and about possibilities for further improvement of the present method.

24/07/07 The future of computational neuroscience

David Willshaw

Components of protein complexes linked to cognition are conserved and co-expressed in the Drosophila brain.

The Wellcome Trust Neuroscience and Mental Health Strategy Group held a meeting earlier this year to identify the exciting new research areas in computational neuroscience. The meeting was centred around a number of talks about specialist areas.

In this talk I will review what the meeting was all about and discuss any consequences of the meeting.

17/07/07 the attraction of Reinforcement Learning

Andrew Gillies

Reinforcement learning was mentioned at the ANC review day. In certain domains of neuroscience and in neuroinformatics it is an area of fast growing interest. This talk will look at why this is happening and some of the key issues open to investigation.

17/07/07 Components of protein complexes linked to cognition are conserved and co-expressed in the Drosophila brain.

Bilal Malik

Neuronal synapses are the fundamental structures that are required for, processing, amplification and strengthening of signals which are manifested in behaviour and disease. The synapse has a large repertoire of proteins most of which interact with each other and are arranged in complexes. In mammals, proteomic studies have focussed on an electron dense zone called the post-synaptic density (PSD), which houses more than 1100 proteins. The PSD hosts the NMDA receptor complex the core of which, in mouse, consists of 186 proteins and is tightly connected to the intracellular signaling machinery and structural proteins via adaptor proteins called Membrane-Associated Guanylate Kinase (MAGUK) proteins. The NMDA receptor complex and the MAGUK proteins together form a complex known as MASC (MAGUK associated signalling complex). Previous studies have shown that MASC proteins co-ordinate signalling to various effectors signalling pathways some of which are crucial for development and maintenance of synaptic plasticity, behavioural processes and in humans are thought to underlie cognitive illnesses. Interestingly, some 89 close orthologues of the MASC proteins are found in the fly, Drosophila melanogaster. Many of these are known to be expressed in the nervous system at some stage during development and a high proportion are linked to behavioural phenotypes. In my brief talk I shall try to give an overview of the complex in the fly and our recent results obtained by Immunohistochemistry and proteomic studies.

03/07/07 Could synaptic depression be on the neural basis of visual adaptation ?

Jesus Cortes

Synaptic depression has been widely studied during the last years. By using a very simple computational model, in this talk I will speak about how synaptic depression affects different properties of cortical cells in the scenario of orientation selectivity. More precise, I will discuss how it affects the tuning curves and the decoding procedure. Finally, I will show how a standard model of orientation selectivity in presence of depressing synapses can explain tilt effects after visual adaptation. I am not saying that visual adaptation is a consequence of synaptic depression, but some properties of visual adaptation emerge for free as a consequence of synaptic depression.

This work has been done in collaboration with Daniele Marinazzo, Peggy Series and Mark van Rossum.

03/07/07 Input-driven oscillations and short term plasticity

Daniele Marinazzo

Synchronization in neural systems can emerge as a result of the architecture or as an effect of an external stimulus. I present a simple two-layer network in which short term plasticity allows to encode, by means of synchronization response, both the temporal characteristics of the input and the presence of spatial correlations. On the other hand, synchronization can give only partial information in the presence of conduction delays and variable synaptic strength. So, a recently proposed nonlinear extension of Granger causality is used to map the dynamics of a neural population onto a graph, whose community structure characterizes the collective behavior of the system. Both the number of communities and the modularity depend on transmission delays and on the learning capacity of the system.

08/05/07 Neural Connectivity in Drosophila melanogaster

Mark Longair

In this talk I will discuss my progress in developing tools to extract neural connectivity information from 3D image stacks of fruit fly brains produced by confocal microscopy. After the image acquisition this process involves registration, semi-automatic tracing and generating readable connectivity graphs from these traces. I will discuss ways of deploying these tools online and plans for future work to improve accuracy of the analysis to the single neuron level and reduce errors from registration step.

08/05/07 Spectral Clustering for Approximate Inference

Felix Agakov

I will discuss an application of spectral clustering to approximate variational inference in undirected models. Commonly, spectral clustering methods are used for clustering the data by utilizing properties of the similarity matrix of the visible patterns. Another application of clustering would be to cluster _variables_ based on the similarities between their marginal and/or conditional distributions. I will discuss possible ways of doing this approximately for fully-connected models. Extracted clusters of variables may then be used for approximate (structured mean field or cluster variation) statistical inference. The purpose of the talk would be to highlight some of the apparently under-explored issues of choosing good clusters for approximate inference in undirected models.

17/04/07 Reversible Jump and an application to a phylogenetic factorial hidden markov model.

Wolfgang Lehrach

In this talk I hope to explain some of the mechanics of Reversible Jump, and an application to phylogenetics:
The traditional approach to phylogenetic inference is based on the assumption that we have one set of hierarchical relationships among the taxa, represented by a binary tree. While this approach is reasonable when applied to most DNA sequence alignments, it can be violated in certain bacteria and viruses due to interspecific recombination, which is a process whereby different strains exchange or transfer DNA subsequences. If undetected, the presence of these so-called mosaic sequences recombinant segments can lead to systematic errors in phylogenetic tree estimation. Their detection, therefore, is a crucial prerequisite for consistently inferring the evolutionary history of a set of DNA sequences. We propose a phylogenetic factorial hidden Markov model that detects these recombinant regions, while also detecting regions with a similar divergence rate and regions with a similar transition-transversion ratio (ttratio). To detect the rate and ttratio regions, we must specify some rates and ttratios which the model switches between along the multiple-sequence alignment. However, instead of using pre-specified rate and ttratio states, the set of divergence rates and ttratios is explored using a reversible jump Markov Chain Monte Carlo scheme. This gives an estimate of the number of different rates and ttratios present in the alignment, as well as estimates of the rate and ttratio along the alignment.

17/04/07 Kernel Multi-task Learning using Task-specific Features (Joint work with Felix Agakov and Chris Williams)

Edwin Bonilla

In this work we address the problem of multi-task learning when task-specific features are available. We describe two ways of achieving this using Gaussian process predictors: in the first method, the data from all tasks is combined into one dataset, making use of the task-specific features. In the second method we train specific predictors for each reference task, and then combine their predictions using a gating network. We demonstrate these methods on a compiler performance prediction problem, where a task is defined as predicting the speed-up obtained when applying a sequence of code transformations to a given program.

10/04/07 A role for Drosophila echinoid in axonal pathfinding and targeting?

Jo Young

The Drosophila cell adhesion molecule Echinoid has been characterised in peripheral nervous system development and has been implicated in neurogenesis through two major pathways. Echinoid is a member of the IgG superfamily and is similar to L1-type molecules which are strongly associated with neurite extension and axon pathfinding (Hortsch 2003) and more recently synaptic function (Godenschwege 2006). Echinoid's role in the developing CNS has not been known or studied to date. Our knowledge of how neurons select specific cells as synaptic targets is limited, by determining which molecules are involved in this process we can start to understand how this is achieved. This study aims to analyse Echinoid's function in the developing Drosophila brain using the photoreceptors in Optic lobe as a model. This project has found extensive Echinoid expression at various points during CNS development, especially between the late 3rd instar and early adult stages in the brain, in addition to mutant targeting phenotypes. This suggests Echinoid may be involved in pathfinding or it could be playing a role in synaptic targeting or formation.

10/04/07 Broca's Area, its Partes and speech

Marina Papoutsi

Broca's area has been a major focus in language (and not only) research for more than a century. It's diversity of function and its peculiarity in anatomy as well as its major role in the "unique among humans" function of speech, have made sure that people will still talk about it. I will briefly talk about some of the features of the region and present some of my related research and initial findings.

20/02/07 Multi-modal segmentation and similarity measures for non-rigid MRI registration

James Withers

The registration of MRI volumes involves finding the best transformations to match the same tissue at each volume pixel (voxel). One recent development in the field is to use segmentation of the brain images into different tissue types in order to assess similarity.

A super-resolution procedure using reverse diffusion to create sharp boundaries between tissues has been used to smooth the images and reduce aliasing as a pre-processing step, with a novel process exploiting structural features of the brain to preserve small detail and not noise. The improvements to segmentation will be evaluated in terms of their effect on the similarity measure over rigid volume movements.

20/02/07 About the Recognizing Textual Entailment challenge

Stefan Harmeling

In order to explore potential applications for machine learning in natural language processing (NLP), I decided to work on the third Recognizing Textual Entailment challenge (RTE, see http://www.pascal- The RTE challenge requires participants to write a program that can decide, whether some short hypothesis (expressed in English) is implied by some longer piece of text. Besides presenting the RTE challenge itself, this talk will introduce some basic notions from NLP that might be relevant for solutions (like part-of-speech tagging, parsing, semantics), useful standard tools (like WordNet), some previous approaches and my (epsilon-baked) plans towards a solution (which might not involve machine learning at their current state).

20th Feb 07

Tim O'Leary

A decade ago spike timing-dependent synaptic plasticity (STDP) was characterised in dissociated hippocampal cultures (Bi & Poo, J Neurosci 1998). I shall present the results of some experiments which go some way toward reproducing this phenomenon and offer some interpretations of these and earlier data. In addition, I shall outline an interesting but technically challenging experiment that aims to dissect the contribution of NMDA receptors to the expression of STDP using a combination of molecular biology and electrophysiology. I will fully welcome suggestions for possible experiments using this system, particularly those with a motivating computational component.

23/01/07 Modelling the development of topographic and laminar organization for orientation and spatial frequency in the early visual pathway

Chris Palmer

Mammalian primary visual cortex contains a variety of topographic maps for visual features, such as orientation and eye preference. Neurons have also been found to have different spatial frequency preferences, but there is widespread disagreement about whether spatial frequency is mapped continuously in cortex, and if so how.

Here, we use a biologically plausible Hebbian-based computational model to make predictions about the possible forms this mapping may take.

We show that matching orientation maps can develop in seperate layers for cortex sheets receiving input from different pathways. This is crucial for reproducing the existing data. By manipulating the strength of interconnections between the laminae, we show that different topographic and laminar SF organizations can develop.

The models make specific predictions that can be tested experimentally to help resolve the controversy surrounding spatial frequecny organisation.

23/01/07 Hierarchical object recognition

Chris Williams

An attractive approach to object recognition is based on a hierarchical scheme, where one first groups low-level features into subcomponents, and these subcomponents are then grouped into components recursively, up to objects. I will give a brief overview of both feedforward and belief network models with this kind of structure. Note that in belief network models there can be top-down influences as well as bottom-up.

16/01/07 Binocularity, depth perception and reading

Richard Shillcock

I will outline what the two eyes do in reading. (They are typically not fixating exactly the same place.) I will then discuss how such behaviour may be related to depth perception in normal close-viewing unconnected to reading, and illustrate some of the implications for learning to read, reading impairment and individual differences.

16/01/07 Revealing neural interaction through the confounds of BOLD

Lawrence Murray

The use of fMRI in the analysis of connectivity between remote brain regions is complicated by the BOLD signal being only an indirect measure of neural activity. While the coupling between neural activation and hemodynamics remains poorly understood, biologically plausible models consistent with experimental results have been developed, and promise to more clearly elucidate the latent neural activity underlying an observed BOLD signal.

In this talk I will explain how connectivity analyses can benefit from the inclusion of such hemodynamic models, with a particular focus on the Balloon-Windkessel model. Existing models that employ such techniques will be discussed, and preliminary results of a novel application using the Unscented Kalman Filter presented.

09/01/07 Reconciling Models of Surround Modulation and V1 Orientation Map Development

Judith Law

Many models of orientation map development have been very successful in reproducing the features of biological maps. The majority of these models are based on a principle of short-range excitatory and long-range inhibitory connections between neurons. However, biological data suggests that long-range connections between V1 neurons arise primarily from putatively excitatory pyramidal cells. Furthermore, simple models with long-range excitation and short-range inhibition have shown how a biologically realistic circuitry can reproduce features of adult V1 function such as extra-classical receptive field phenomena (Schwabe et al., 2006). Previous developmental map models with long-range inhibitory connections are therefore unable to account for aspects of surround modulation. I will be presenting my Phd work so far, which has been investigating how such circuits can arise, which parts of the system are plastic, and in general how to reconcile these findings with otherwise successful developmental models such as LISSOM.

12/12/06 Chasing the Homonculus: Decoding populations of neurons after visual adaptation

Peggy Series

I am going to present a project I am currently working on, in collaboration with Eero Simoncelli, at NYU.
Electrophysiological data show that prolonged exposure to visual patterns profoundly modifies the responses of visual neurons: neurons adapt and their responses usually decrease. In psychophysics, the same visual adaptation protocol results in biases in the estimation of the parameters of the stimuli (eg. the tilt after-effect), visual illusions (eg. the waterfall illusion), as well as changes in discrimination performances. This project aims at understanding the relationship between the physiological effects of adaptation and their perceptual consequences. We model the effect of adaptation at the level of a population of neurons (eg. in V1). We then investigate how the neuronal responses can be "decoded", using different types of estimators, eg. maximum likelihood, population vector, winner-take-all etc... We compare the prediction of these estimators with perceptual data, in terms of both orientation estimation (bias) and orientation discrimination performances (variance).

05/12/06 Predicting microRNAs in thalamocortical axons

John Davey

I will discuss my work on brain development, specifically the growth of axons from the thalamus to the cortex. I am studying a set of messenger RNA (mRNA) molecules we believe are present in these axons. I will introduce a set of small RNA molecules called microRNAs that are known to regulate mRNA molecules in developing cells and have been implicated in the onset of cancer. I am trying to discover whether particular microRNAs target the mRNAs we have found in thalamocortical axons. This is a problem which can be approached bioinformatically. I will describe the basic algorithms used for attacking the microRNA target prediction problem.

05/12/06 Memory' Stability and Spike Timing Dependent Plasticity

Guy Billings

We all know that by using learning rules and neural networks we can create systems that learn statistical properties of the data. But what happens to the weights after that? If the system is trained to perform a task and then the weights are frozen, the answer is 'not much'. But in general biology does not freeze weights after training. This raises the question of how biological systems might retain memories despite constant weight perturbations. As some of you already know, I have been examining two models of Spike Timing Dependent Plasticity - an important but most likely flawed plasticity paradigm - that are both supported by experimental evidence. In particular I have been asking how long learned weight correlations are retained given fluctuating neural activity levels. That is to say: How stable are the 'memories'? In this talk I will present the latest developments in my PhD research.

28/11/06 What's up with plasticity?

Mark van Rossum

So you thought all there is to plasticity is dw = x.y ? In this journal club style talk I will review some recent data on the physiology of LTP/LTD how long it lasts, and synaptic tagging. I will hint towards possible unifying models.

21st Nov 06

John Quinn

Kalman filters are 46 years old and still very influential in time series modelling. However, their training and verification are sometimes regarded as something of an art, which might be off putting to the casual user. I'll review some training techniques and focus on ways of verifying that the model is right, illustrated with some different types of data and the 'Kalman Filter Diagnostic Dashboard', a simple tool.

21/11/06 21st Nov

Adriano Werhli

An important problem in systems biology is the inference of biochemical pathways and regulatory networks from postgenomic data. The availability of high-throughput postgenomic data has recently prompted substantial interest in reverse engineering the networks and pathways in an inferential way from the data themselves. Microarray experiments are very popular nowadays. While they can measure gene expression in thousands of genes at the same time, this measurements still very sparse and noisy. This makes the reverse engineering of the networks and pathways based solely in microarray data very unreliable. In this talk I will present a method for integrating other sources of knowledge when reverse engineering networks and pathways. These extra sources of knowledge can come from other experiments, literature information or any source of knowledge about the system being studied. We call it prior biological knowledge, and it is combined with the microarray data during the inference. I will show some examples how the algorithm works using simulated data and some preliminar results we obtained when applying the method to real data sets.

31/10/06 Experimental and computational investigations into the rodent head direction cell system: a 2-part rapid-fire talk on my thesis work

Matthijs van der Meer

Part 1: Neural representations of place and direction, like the grid and head direction (HD) cells found in the entorhinal cortex and associated areas in rodents, are thought to be critical for various spatial behaviours. However, experimental evidence for such a role has generally been indirect and even conflicting. Our approach is to study path integration, the updating of one's position using internal information only, recording from HD cells as rats perform a spatial task which we show does not depend on external cues. Combined with a lesion study on the same task, the results provide evidence that when path integrating, rats rely on their HD cells for spatial orientation.

Part 2: What is the neural mechanism underlying the maintenance and updating of the HD signal? HD cells can anticipate the animal's HD by up to 75 ms, but current models cannot account for any of the substantial variability in this anticipation. We consider how the statistics of rat head movements (as extracted from a large corpus of data from a number of different behavioral tasks) interact with putative neural mechanisms, based on the firing properties of neurons afferent to the HD system. In direct comparison with experimental data, we can explain up to 50% of the experimentally observed variability in this way.

07/11/06 Generation of artificial spike trains

Irina Erchova

The Poisson process is often taken to be a good starting place for modelling spike trains. This assumption ignores a refractory period (that is easy to deal with) and a relatively long period of after spike hyperpolarisation (AHP) when probability of spike generation is reduced. As a result, a number of data analyses techniques (like reverse correlation and pattern detection) suffer from biases. I am going to discuss two alternative approaches: integrate (or resonate) and fire model (that can be treated analytically in some cases) and a gamma process.

31/10/06 A novel route for the pathophysiological cascade resulting from dopamine depletion.

Andrew Gillies

The subthalamic nucleus (STN) is the primary target for surgical treatments of Parkinson's disease, such as chronic deep brain stimulation. However, the role of the STN in Parkinson's disease, and the mechanisms by which the surgical interventions alleviate Parkinsonian symptoms, remain poorly understood. In rat, STN projection neurones exhibit physiological characteristics that share features with the aberrant activity observed in animal models of Parkinson's disease. In vitro, hyperpolarisation of STN neurones can lead to the emergence of rhythmic bursting and initiate variable length rebound bursts.

Computational models, able to reproduce the key STN physiology, lead to the prediction that there is a mechanism located in the dendrites able to generate depolarising plateaux. A tight coupling of the L-type and T-type calcium channels (CaV 1.2-1.3 and CaV 3-respectively) provides a mechanism for prolonged depolarisations,facilitated through hyperpolarisation. In addition, the small current calcium activated potassium channel (KCa 2.1-2) modulates the resulting plateau. As the L-type channel is critical to both the rhythmic bursting and rebound behaviours in the model, this provides an intriguing possibility for direct dopaminergic modulation of this channel and influence over the emergence of these behaviours.

24th October 06

Douglas Armstrong

Proteomic study of the mammalian synapse has generated an extensive list of molecular components, revealing it as one of the most complex biological systems. While fundamental to information processing, behaviour and cognitive disorders, the molecular architecture of signalling in the synapse and its relation to higher-level function is now beginning to emerge. We present a model of the synaptic proteome that captures its structural organisation. Each component is annotated with information describing its molecular features/domains, evolutionary orthologues across 19 species, measurements of gene/protein expression in brain regions and functional annotation from yeast, fly, mouse and human. The model reveals a highly integrated and modular structure. Modules, defined by molecular interactions, not only share common network properties but also functional annotation, regional expression patterns and evolutionary origins. The picture that emerges is of a set of input modules (e.g. receptor complexes) that are closely linked to higher order cognition (and disorders), of more recent evolutionary origin and with high regional variation in the brain. These then link through more central processing modules to a series of output modules (e.g. gene regulatory complexes) that are more closely related to vital functions, ancient evolutionary origin and little variability across the brain.

26/09/06 TIPs and High-Level Models for Activity Recognition

Tom Griffiths

In this talk I'll outline my progress on an object-based approach to the computer vision task of human activity recognition from video. Still in its relatively early stages, I'll show progress on object tracking, temporal interest point (TIP) extraction, and share some thoughts on high-level models. I'll also introduce a new dataset I've been working on.

26/09/06 Learning with Transference: an Overview

Felix Agakov

In this talk I will review some of the common and more recent approaches to transfer learning, i.e. improving performance on a new problem using what has already been learned about other "similar" problems. I will summarize and discuss some of the more general methods presented at the NIPS 2005 workshop "Inductive Transfer : 10 Years Later". I will also discuss some of the supposedly new ideas on transfer learning and demonstrate some results for an applied problem of evaluating the performance of compiler optimizations.

19/09/06 How to be forgetful: part II

David Sterratt

My previous work with David Willshaw on forgetting in linear associative networks has considered exponential forgetting, where the intensity of a memory trace in the network decays exponentially with its age. Prompted by work by Stefan Fusi and Larry Abbott and conversations with Jesus Cortes, in this talk I will consider what happens if the strength of a memory decays according to a power law. I will show that exponential forgetting can always be optimised to outperform power-law forgetting, but if Nature is not allowed to tune the forgetting rate, power-law forgetting is superior.

19/09/06 A Phylogenetic Reversible Jump Markov Chain Monte Carlo Factorical Hidden Markov Model

Wolfgang Lehrach

The traditional approach to phylogenetic inference is based on the assumption that we have one set of hierarchical relationships among the taxa, represented by a binary tree. While this approach is reasonable when applied to most DNA sequence alignments, it can be violated in certain bacteria and viruses due to interspecific recombination, which is a process whereby different strains exchange or transfer DNA subsequences. If undetected, the presence of these so-called mosaic sequences recombinant segments can lead to systematic errors in phylogenetic tree estimation. Their detection, therefore, is a crucial prerequisite for consistently inferring the evolutionary history of a set of DNA sequences. We propose a phylogenetic factorial hidden Markov model that detects these recombinant regions, while also detecting regions with a similar divergence rate and regions with a similar transition-transversion ratio (ttratio). To detect the rate and ttratio regions, we must specify some rates and ttratios which the modelswitches between along the multiple-sequence alignment. However, instead of using pre-specified rate and ttratio states, the set of divergence rates and ttratios is explored using a reversible jump Markov Chain Monte Carlo scheme. This gives an estimate of the number of different rates and ttratios present in the alignment, as well as estimates of the rate and ttratio along the alignment. I'll cover some of the difficulties with implementing this approach, and its application to some real examples.

12/09/06 Measuring Maps

David Willshaw

There are at least two candidate mechanisms, each supported by a wealth of empirical data, for how somatotopic maps of connections are formed in the nervous system. These are: an activity-based mechanism and a molecular-based mechanism; most likely, both mechanisms are involved in map-making. Advances in experimental techniques for disturbing the putative map-making mechanisms in a controlled fashion to test out particular theories and techniques for assessing the pattern of connections formed now make it possible for the contributions of different mechanisms to be assessed.

A prerequisite for this is a method for measuring the precision of a map. Up till now, no such method has existed. I will describe the method that I have been developing by which (a) the precision and (b) the orientation and polarity of a map can be measured. I will compare this with other possible approaches, such as those that could be based on the Procrustes method for matching together two sets of points embedded in a multidimensional space.

05/09/06 Modelling handwriting as a generative process

Ben Williams

Using a probabilistic framework, it is possible to extract motor primitives from handwriting data. These primitives can reconstruct the data, and are not limited to any particular section of the character, or data set. The model can be run as a generative process, which will produce samples of handwriting. A complete generative model needs a timing model for the primitive activation likelihoods, the absence of which produces a scribbling form of output. Different timing models are possible, and implementations of which have performed with varying success. I shall present some background to this research, the primitive and timing models, and the impact that the different timing models have upon the handwriting output.

05/09/06 Predicting the Performance of Compiler Optimizations using Transference (Joint work with Chris Williams and Felix Agakov)

Edwin Bonilla

Given a program, we wish to make it run fast by finding a code transformation sequence that optimizes its running time. Compiling and running the code to explore this space of transformations can take a very long time, so we investigate the use of predictors (or proxies) that take as input a program X and a transformation sequence t, and predict the speed-up of that combination. Our method is not specific to a given program, but transfers information across programs and so can be applied to a new program we have not seen before. We have found that response-based features are more effective than code-based features for carrying out these predictions. Our method is demonstrated on the UTDSP suite of benchmarks.

29/08/06 Likelihood-based estimators in the example of Robins-Ritov (joint work with Marc Toussaint)

Stefan Harmeling

Bayesian or likelihood-based (LB) approaches to data analysis became very popular in the field of Machine Learning. However, Robins and Ritov (1997) showed that LB estimators are lacking certain possibly desirable frequentist properties by constructing a special example in which (generic) LB estimators will fail. This questions the general applicability of the LB method. In this talk we consider various approaches to formulate LB estimators for their example and for illustrative simplifications thereof. Basically, we conside extensions of the presumed generative model of the data. Hereby, we can derive estimators which are very similar to the classical Horwitz- Thompson estimator and which also account for a priori knowledge of an observation probability function. Note that we are not disproving Robins and Ritov's result but we try to shed some light on its practical impact.

29/08/06 When Training and Test Distributions are Different

Amos Storkey

In reality we all build models on the basis of information obtained in one domain, but fully intend to apply them to a different domain with, if we were honest, slightly different characteristics. This goes for neuroscientists, cognitive scientists and bioinformaticians as much as it does for machine learners. What should we do? Just hope for the best?

In fact statisticians have spent some time characterising common situations where training and test distributions are different, and some effort working out how to deal with them. Even so there is room for further development in this area, and it is currently seeing a revival of interest.

In this talk I will outline the common cases for changing environments, and then discuss the basic ideas behind an approach to one case we think is of critical importance: regression problems where a single dataset consists of data from a number of sources, and the proportions of these sources vary across datasets. I will also relate it to other work on topic modelling.

This is joint work with Masashi Sugiyama

PS I will also be advertising the art of Elin Bjorsvik, who is currently exhibiting at the first floor of the C Fringe Venue on Chambers Street. However the exhibition will be over by the time I give the talk, so do give yourself some time to take a look if you are popping to some fringe event there, especially if you are blue/red depth illusion sensitive!

21/08/06 Eye movements in reading development

Richard Shillcock

I will run through some of what is known about how eye movements change during reading development (i.e. through primary and secondary education). I will address the question of whether these changes reflect higher-order, comprehension processes or lower-level word recognition processes, and present some modelling data suggesting that the latter are involved. In general, I will talk about the nature of causal explanation in understanding the eye movements that go on during reading.

21/08/06 Machine Learning and Statsitics: So What's the Difference?

Chris Williams

The construction of unconditional or conditional probabilistic models for data is widespread across the fields of statistics and machine learning. I will provide some examples of the rich and varied kinds of data and models used in machine learning, and also discuss the differences in motivation and emphasis between the two fields.

15/08/06 Neuroanatomy of the Central Complex in Drosophila

Mark Longair

In this talk I will describe the progress of my work towards building a quantitative atlas of the central complex of Drosophila melanogaster, including acquiring the data, evaluating image registration algorithms, building template images and the future direction of my PhD.

08/08/06 Recognising objects using a Generative Template of Features

Moray Allan

I will discuss some results on recognising object categories in cluttered images. We represent images as collections of 'visual words', by clustering descriptors computed for regions around interest points. We model visual word distributions and locations using a Generative Template of Features (GTF), consisting of a number of 'parts', where each part has a corresponding distribution over visual words.

08/08/06 8th August

Jo Young

The Drosophila cell adhesion molecule Echinoid has been characterised in peripheral nervous system development and has been implicated in neurogenesis through two major pathways. Echinoid is a member of the IgG superfamily and is similar to L1-type molecules which are strongly associated with neurite extension and axon pathfinding and more recently synaptic function. Echinoid's role in the developing CNS has not been known or studied to date. This project has found extensive Echinoid expression at various points during CNS development, especially between the late 3rd instar larva and early adult stages in the brain. This may indicate that the molecule is involved in axon pathfinding or synaptic targeting. In order to study the role of Echinoid in more detail, it is necessary to use a well characterised system such as the optic lobe. This talk will highlight recent work on the project and its future direction.


Guy Billings

Synaptic plasticity is thought to underpin learning and memory. This hypothesis supposes that activity within the brain, in the form of neuronal action potentials, changes synaptic efficacies or ’¡Æweights’¡Ç. Synapses alter how strongly the activity of a particular neuron is coupled to the activity of other neurons. Consequently changes to synaptic efficacies can in turn lead to changes in future neuronal activity. According to this picture memories are made of interactions between experience and physical alteration of synaptic efficacies.

We know that memories can last for long periods of time within the brain. If this is the case then it is likely that there must be groups of synapses whose efficacies are stable over long periods of time. However the synapses encoding this information are occasionally activated by new learning, random activity or reactivation of the old memories. In this case their efficacies can be modified leading to a deletion of the stored information. In light of this we may ask: How is information stored for long periods in the brain? In this talk I will briefly give an overview of some previous work that illustrates this problem in the case of a specific model of synaptic plasticity; Spike Timing Dependent Plasticity. I will then introduce some developing work that aims to tackle this question but in the more general framework of a Markov model of the synapse.

01/08/06 1st August

Jon Claydon

Methods for visualising the brain's white matter have begun to be used as tools for selectively studying the effects of pathology on specific structures. While a considerable amount of work is being put into improving the effectiveness of tractography algorithms for doing the "selection" of a structure of interest, most studies so far have simply averaged diffusion anisotropy - a proxy for white matter integrity - in the segmented regions. Such averaging reduces noise issues, but provides very limited information from which to draw conclusions. In this work, we develop a technique for studying the anisotropy profile along a specific white matter tract for groups of subjects, and apply it to compare data from vascular cognitive impairment patients with data from normal volunteers.

25th July 06

John Quinn

Someone wants to sell you a model for some sequential data (e.g. horse racing results, stocks and shares, or lottery numbers), claiming that you will profit if you adopt it. Should you spend any money on it? The first part of this talk will be a general introduction to deciding whether a dynamic model fits the data it is modelling, illustrated using Kalman filters. I will then relate this to recent work on physiological monitoring of premature babies, in which we want to try and guess the periods in which unrecognised (and possibly dangerous) types of variation are occurring, by assessing the plausibility of the 'stable baby' model.

25/07/06 25th July

Jane Ewins

Brainwave is a unique assay that allows users to characterize compounds interacting with molecules and cell-signalling pathways in the nervous system. New chemistries with a specific mechanism targeting neurotransmitter pathways are needed and highly valued. Pharmaceutical trends show fewer new products being supplied to the marketplace, but increased demand for new treatments for the most prevalent CNS disorders: pain, depression, epilepsy, schizophrenia, migraine, Alzheimers and Parkinson?s disease. Brainwave is a bioassay technology which complements smarter, lower-risk product development strategies by choosing compounds that are most likely to succeed, and reducing costly drug failures in pre-clinical and clinical trials. In this presentation I will be presenting the technical development of this technology, demonstrating the sensitivity of the assay to specific compounds, and the discussing future applications of using Drosophia (fruit fly) in CNS drug discovery and development.

11/07/06 A competition among depression, facilitation and noise in recurrent networks with oscillatory neural activity

Jesus Cortes

In visual cortex depressing synapses are dominant. Contrarily, in medial prefrontal cortex Markram et al in [1] have found by using multineuron patch-clamp recording in slices of young ferrets a high heterogeneous population with different kind of synapses. In agreement these results and computer simulations, we model (for details see [2]) recurrent networks in presence of both depressing and facilitating synapses and noise. As a result, our model is a simplified tool to understand the underlying of different competing synaptic mechanisms and their functionality in working memory within sustained cortical activity.

[1] Y. Wang, H. Markram, P.H. Goodman, T.K. Berger, J. Ma and P.S. Godman-Rakic. Heterogeneity in the pyramidal network of the medial prefrontal cortex. Nat Neurosci. 9, 534-542 (2006)

[2] J.M. Cortes et al. download at

27/06/06 Seconday results from whole-genome, RIDGE, analysis of microarray data

Adriano Werhli

Comparative Evaluation of the Accuracy of Reverse Engineering Gene Regulatory Networks With Various Machine Learning Methods

An important problem in systems biology is the inference of biochemical pathways and regulatory networks from postgenomic data. Various reverse engineering methods have been proposed in the literature, and it is important to understand their relative merits and shortcomings. In the present paper, we compare the accuracy of reconstructing gene regulatory networks with three different modelling and inference paradigms:
(1) Relevance networks (RNs)
(2) graphical Gaussian models (GGMs):
(3) Bayesian networks (BNs):
The evaluation is carried out on the Raf pathway, a cellular signalling network describing the interaction of eleven phosphorylated proteins and phospholipids in human immune system cells. We use both laboratory data from cytometry experiments as well as data simulated from the gold-standard network. We also compare passive observations with active interventions.

20/06/06 Seconday results from whole-genome, RIDGE, analysis of microarray data

Lena Hansson

20/06/06 20th June 2006

Michael Schouten

his talk will focus on the causes for high level of Type I error rates in genetic association studies. I wil show that using graphical models to account for dependencies in data can reduce these rates.

13/06/06 Mechanisms of rhythmic bursting in rat subthalamic nucleus projection neurons

Andrew Gillies

The rat subthalamic nucleus (STN) projection neurons exhibit a number of characteristic bursting properties under hyperpolarising conditions. Variable length post-hyperpolarising rebound bursts are observed across the STN population. In the presence of apamin, constant hyperpolarisation can lead to slow rhythmic bursting. Here I present a set of simplifications to our recent full compartmental model, systematically reducing the number of compartments and the channel composition to isolate the key anatomical and channel arrangements that reproduce the characteristic bursting properties. This shows that the interactions of the L-type and T-type calcium channel populations (Cav 1.2-1.3 and Cav 3- respectively) with the small current calcium activated potassium channel (Kca 2.1-2) underlie these behaviours. The co-localisation of the L-- and T-- type channels mediates the generation of a plateau potential that facilitates extended bursts. Reducing the model to a single compartment cannot reproduce the phenomena, implying a dendritic location for the plateau and bursting mechanisms.

13/06/06 13th June 2006

Michael Schouten


06/06/06 Latencies in the visual system

Mark van Rossum

In this talk I will show how Monte Carlo simulation techniques can be used to investigate problems in neuroscience. These methods are particularly useful to study processes on very small spatial and temporal scales, where the relevant variables are typically experimentally no longer directly accessible. I will then introduce the MCell simulation package, which uses highly optimised Monte Carlo techniques to simulate cellular micro-environments. Finally I will show results of Monte Carlo simulations of synaptic transmission at the Calyx of Held, a giant glutamatergic synaptic terminal in the mammalian auditory brainstem. This work was carried out last year in Stirling (with Bruce Graham and the group of Ian Forsythe in Leicester) as part of a project to develop detailed models of excitatory synaptic transmission.

06/06/06 Microphysiological models in neuroscience - Monte Carlo simulation of cellular micro-environments

Matthias Hennig

In this talk I will show how Monte Carlo simulation techniques can be used to investigate problems in neuroscience. These methods are particularly useful to study processes on very small spatial and temporal scales, where the relevant variables are typically experimentally no longer directly accessible. I will then introduce the MCell simulation package, which uses highly optimised Monte Carlo techniques to simulate cellular micro-environments. Finally I will show results of Monte Carlo simulations of synaptic transmission at the Calyx of Held, a giant glutamatergic synaptic terminal in the mammalian auditory brainstem. This work was carried out last year in Stirling (with Bruce Graham and the group of Ian Forsythe in Leicester) as part of a project to develop detailed models of excitatory synaptic transmission.

30/05/06 Image processing techniques for automating malaria diagnosis

Stephen Henderson

Malaria is a parasitic disease which infects red blood cells. Infected cells can be classified to give the specific strain and stage of infection by expert examination of blood smear samples under a microscope. I will discuss some potential techniques for automating the diagnosis procedure via image processing software as well as some of the issues involved in dealing with these types of images.

30/05/06 Robust performance of inhomogeneous forgetful associative memory networks

David Sterratt

(in collaboration with David Willshaw)
We investigate how various inhomogeneities present in synapses and neurons affect the performance of the linear associative memory model, a high level model of hippocampal circuitry and plasticity. The inhomogeneities incorporated into the network model are: differential input attenuation, stochastic synaptic transmission and memories learnt with varying intensity. We determine the memory capacity of the network with an arbitrary learning rule using a the signal to noise ratio (SNR) analysis. The SNR in an inhomogeneous network is related to the SNR in an equivalent homogeneous network by a factor that depends on the coefficients of variation (CVs) of the attenuation factors, stochastic transmission factors, and the learning intensity of the memories.

We apply our expression to the attenuation due to extended dendritic trees by choosing distributions of attenuations that would result branched dendritic trees. We use biological parameters for stochastic transmission to determined the CV of the transmission factors. We find that the reduction in SNR due to differential attenuation is surprisingly low compared to the reduction in SNR due to stochastic transmission. This is perhaps surprising in the context of the findings that synaptic conductances are scaled strength according to distance from the soma. Nevertheless, our results also show that the effects of compensating for attenuation will not be diminished by other inhomogeneities, due to the factorial combination of the changes due to each type of inhomogeneity.

We apply the general result for patterns learnt with different intensities to a learning rule that incorporates weight decay. This network has the palimpsest property of being able to store patterns continuously at the expense of older ones being forgotten. We show that there is an optimal rate of weight decay that maximises the capacity of the network, and that the capacity of the network is a factor of $e$ lower than its non-palimpsest equivalent. This result has not been derived in the context of other inhomogeneities before.


Tom Griffiths

Human activity recognition is the computer vision endeavour of automatically inferring human tasks and behaviour from video streams. The ability to do this well has many real-world applications. In this talk I will outline our approach to this problem, which attempts to go beyond the use of simple position and body shape information for inference by utilising information about the objects used in tasks. I will describe the three tiers of the problem, and present details of our initial attempt at solving the first: object tracking. Some future directions will then be discussed.

09/05/06 Chop stick snapping in Chinese Restaurants

Stefan Harmeling

After a brief recapitulation of what is Bayesian inference, a brief introduction to Dirichlet processes will be given, hereby employing the usual vivid analogies of Chinese Restaurant processes, stick-breaking processes and all that. If time allows, Bernoulli trips and Polya trees will be presented as well.

09/05/06 Nonlinear dimensionality reduction by maximizing information content

Felix Agakov

In the first part of the talk I will outline some problems of applying the infomax principle for visualization and manifold learning. Then I will show that such problems may potentially be alleviated by carefully choosing constraints on the projections to the lower-dimensional space. I will draw a link between the constrained infomax formulation and some of the common spectral methods for nonlinear dimensionality reduction.

If time permits, I will also outline: (a) a simple way of applying the information-maximization principle for extracting a subset of patterns which are predictive about a specific regression problem (which may be used for learning with transference); (b) a simple way of applying the variational EM for conditional learning of population codes (following Peter Latham's ANC seminar given on Apr. 4).

02/05/06 Models for the elimination of superinnervation of the developing muscle and their relation to Motor Neuron Disease and related conditions

David Willshaw

During the development of the nervous system, muscle fibres are initially innervated by more than one motor axon. During the first few weeks of life, axonal contacts are withdrawn to leave each fibre innervated by a single axon. The same process takes place during the reinnervation of muscle following injuries.

I will review computational models for the withdrawal of superinnervation, developed by myself, Carl Rasmussen and Arjen van Ooyen in three separate pieces of work.

In many different diseases, notably Amyotrophic Lateral Sclerosis (Motor Neuron Disease), and related animal models, loss of neuromuscular junctions leaves muscle fibres denervated. The normal compensatory mechanism of sprouting of axons onto deinnervated muscle fibres is inadequate to save the progressive lack of function that results. I will explain - tentatively - how one of our models, the Dual Constraint Model, could be applied to model such diseases.


Jim Bednar

he neocortex of mammals is primarily organized into topographic maps, and understanding how these maps develop and function is crucial for understanding the brain. The Topographica neural map simulator is designed to make large-scale maps practical to model computationally, so that the sparse experimental data available can be synthesized into a coherent theory. The simulator is designed to be as general as possible, to avoid putting constraints on the form of models yet to be developed, but it also must support a specific class of models for it to be useful. In this talk, I will discuss the types of models that are supported, the current division into component parts, and how these parts interact. The goal is to get feedback about important models that this framework cannot easily support, ideas for how to improve the component names and duties, and other ideas for future directions for the simulator.

25/04/06 25th April

Wolfgang Lehrach

Peptide recognition modules (PRMs) are specialised compact protein domains that mediate many important protein-protein interactions. They are responsible for the assembly of critical macromolecular complexes and biochemical pathways [PawsonTony1997]), and they have been implicated in carcinogenesis and various other human diseases [Sudel2000]. PRMs recognise and bind to peptide ligands that contain a specific structural motif. One of the most activelystudied PRMs is the SH3 domain, which binds to peptide ligands that contain a particular proline-rich core.

We propose an alternative in silico method for the prediction of SH3-mediated protein-protein interactions, which addresses some of the shortcomings of the model introduced by [Reiss2004]. A key feature of our model is that it is discriminative: given a set of protein sequences, the model only attempts to find domains that distinguish between different SH3 binding domains.

18/04/06 Binocularity in reading

Richard Shillcock

I will give a tutorial introduction to issues of binocular control of eye movements, drawing on the literature from binocular rivalry and eye-tracking and on current experiments. I will be particularly concerned with the implications for reading. I will look at some of the alternatives for combining information from the two eyes and some of the factors about orthographic processing that may make reading a special case.

10/04/06 Extracting Motor Primitives from Handwriting Data

Ben Williams

For the past 10 years it has become clear that biological movement is made up of sub-routine type blocks, or motor primitives, with a central controller timing the activation of these blocks, creating synergies of muscle activation. We have shown that it is possible to use a factorial hidden Markov model to infer primitives in handwriting data. These primitives are not predefined in terms of location of occurrence within the handwriting, and they are not limited or defined by a particular character set. Also, the variation in the data can to a large extent be explained by timing variation in the triggering of the primitives. Once an appropriate set of primitives has been inferred, the characters can be represented as a set of timings of primitive activations, along with variances, giving a very compact representation of the character. Separating the motor system into a motor primitive part, and a timing control gives us a possible insight into how we might create scribbles on paper.

28/03/06 Beyond ICA

Chris Williams

Independent Components Analysis has been used to model, for example, patches of natural images. However, analysis of the resulting latent variables extracted shows that they are, in fact, not independent. In this talk I will describe a model by Karklin and Lewicki (2003, 2005) which seeks to model these correlations, and describe its relationship to the subspace ICA model (Hyvarinen and Hoyer, 2000) and the topographic ICA model (Hyvarinen, Hoyer and Inki, 2001).

28/03/06 Learning state representations with probabilistic inference for solving POMDPs

Marc Toussaint

Solving POMDPs requires to maintain some representation of state, based on which actions can be selected. Standard approaches, such as belief states, history-based and predictive state representations, pre-determine the type of state representation an agent will utilize independent of a specific goal or the behavioral relevance of the information they encode. Our aim is to learn representations for a specific problem or class of problems.

In this paper, we show how internal state representations can be learned using standard probabilistic inference methods: For this we augment the Dynamic Bayesian Network formulation of the POMDP with latent variables that describe the agent's internal state. We show that likelihood maximization in our model is equivalent to maximization of the expected future reward and thus we can use the expectation-maximization (EM) procedure to learn the states' update rules and their coupling to actions. Such a learned internal state representation integrates information gained from past observations that is relevant for goal-directed behavior and neglects behaviorally-irrelevant information. We demonstrate the approach on some maze problems, where internal representations for certain junctions or aisle following reactive behaviors are learned. The framework is in principle extendible to arbitrary latent structures such as factorial or hierarchical internal representations.

28th Feb 06

Jon Claydon

Diffusion magnetic resonance imaging (dMRI) provides information about the nature of connectivity in the living human brain. A standard dMRI brain volume data set provides local white matter directionality information down to a resolution of around 2x2x2 mm. A range of so called "tractography" algorithms have been developed to integrate this information and visualise whole white matter fibre bundles ("tracts"), but it is important for downstream applications that it becomes possible to robustly segment a particular tract in a group of volumes acquired from different subjects. In this talk I will describe a general approach to this reproducibility problem based on algorithmic comparison of the topology of a candidate tract with a predefined reference tract, and outline work that suggests that effective segregation between good and bad matches should be feasible. In addition, I will describe a proposed Expectation Maximisation technique for reproducible tract segmentation, and discuss possible low level models for the dMRI data.

28/02/06 A single neuronal network model of familiarity- and recollection-based recognition memory

Andrea Greve

We are able to store and retrieve different types of information from our memory. Previous research has shown that memory is not holistic but divided into several subsystems supported by different kinds of encoding and retrieval processes. For example, two distinct retrieval processes are recall and recognition. Whilst recall allows us to generate a previously-studied item, recognition describes the ability to identify a presented item as previously encountered. Recognition itself is contingent upon familiarity (a fast acting process that reflects a quantitative assessment of memory strength) and recollection (retrieval of qualitative contextual information). Detailed accounts of how recognition memory is supported by familiarity and recollection have been proposed by Dual-process theories (for review see Yonelinas, 2002).

In this talk I will present a simple neuronal model capable of simulating memory retrieval in line with Dual-process theories. Both, familiarity and recollection discrimination is achieved within a single attractor neuronal network. The present model incorporates the finding that there is a difference between the speed of familiarity and recollection, and that the two processes are functionally independent.

14/02/06 Stability in Spike Timing Dependent Networks

Guy Billings

Changes to synaptic weights that are dependent upon the difference between post-synaptic and pre-synaptic spike times, have been experimentally observed in vitro. These observations raise the possibility that synapses can be modified according to the temporal correlations between a small number of spikes. In contrast, classical plasticity investigations have concentrated on changes associated with large volleys of spikes (they have been rate-based).

The data suggest that the magnitude of synaptic modification can depend upon the current value of the synaptic weight, in addition to the timing differences between pre and post-synaptic spikes for that synapse. However it is still not clear that this is always the case. In other words, the STDP learning rules could be weight dependent or weight independent. Phenomenological models have been proposed that come in both weight dependent and weight independent flavours. For this reason there has been recent interest in understanding the theoretical - and ultimately biological - implications of the weight dependency of STDP learning rules.

In this work I compare a weight dependent and a weight independent model of STDP. Specifically I explore the stability characteristics of the models. Firstly in the case of an ensemble of identical synapses connected to a single integrate and fire unit. Next I consider the case of a recurrent network that has been trained to perform a task. Weight dependency strongly effects the stability of the learning rules and consequently the dynamic properties of networks. I find that this has clear functional implications. Hence this work suggests additional experimental measures that could be used to determine the weight dependency of biological STDP.

14/02/06 14th Feb

Joanna Young

Notch signalling is essential for development of the nervous system in most multicellular organisms and has been implicated in adult brain function. Abnormalities in Notch signalling results in a range of human diseases such as Alzheimer's disease and it has been implicated in cancer. Furthering our understanding of the components involved in this pathway using Drosophila will enable us to combat such diseases more effectively. The CAM Echinoid has been implicated in both EGFR and Notch signalling in the developing fly. Previous studies have associated Echinoid with Neuroglian, Notch and DE-Cadherin, as well as observing homophilic binding. However, the exact role of Echinoid in relation to these other molecules has not yet been deduced.Characterisation of Echinoid's molecular interactions at both the protein and genetic level will improve our knowledge of cell signalling and may add to the current Notch signalling pathway.

24/01/06 Are genes organised in a random fashion in the genome or not?

Lena Hansson

There are various models for gene organisation within the genome. One of these suggests that there are islands, or ridges, of genes that are arranged in a sequential manner along the DNA, and that are co-expressed as a group. Drawing upon evidence from a series of microarray experiments; we look at the gene structure of a specific region, the MHC locus, on chromosome 17 in mouse.

17/01/06 Genes 2 Cognition - an update

Douglas Armstrong

I up present a summary of work done on the Genes2Cognition programme to date and an overview of what we are currently working towards. This will focus on our developing models of the synaptic proteome.


Edwin Bonilla

One interesting scenario in compiler optimisation is iterative compilation, where one can afford several program executions in order to determine a set of program transformations that significantly increase performance. This task can be formulated as a (discrete) combinatorial optimisation problem where a set of transformations can be combined into sequences of arbitrary length. The goal is to find a sequence that minimises an objective function such as the execution time. Instead of designing new search algorithms, it is clear that existing algorithms can be improved by using a search distribution of good solutions that can be learnt from previous experience with other programs. This technique can be referred to as Predicting Search Distributions where one is interested in learning the mapping from program features to a search distribution of transformation sequences. This approach has been applied to a set of C programs in two different embedded architectures, and the results indicate that the learnt models can speed up iterative search on large spaces by an order of magnitude. Going beyond compiler optimisation, the idea of enhancing search by using predicted distributions could be applied to other difficult search problems provided one can find a good characterisation of the domain.


Andrew Gillies

The question of what the basal ganglia do was thought to be solved in the early 1900s by its observable role in movement disorders. It was considered to be involved in voluntary movement control and planning. By the latter half of that century this vague idea was being questioned with increasing physiological data not readily consistant with this function. There are currently two main competing ideas for a general function for the basal ganglia: "reinforcement learning" and "action selection". It is also widely considered that they may be two views of a single function. However, in this talk I want to re-introduce a less considered idea originally proposed for the basal ganglia in the 1990s, and how this forgotten idea may be an important piece to the puzzle.


Matthjs van der Meer

Head-direction (HD) cells in the rodent limbic system can anticipate the animal's head direction by up to 80ms even during passive movement, but the mechanisms underlying this effect are poorly understood. How do they do it, and is it useful for anything? I'll summarise briefly some issues with the current 'offset connections' hypothesis and previously presented work on simulating spike rate adaptation (SRA) and post-inhibitory rebound firing (PRF) in the vestibular inputs to the HD system as an alternative explanation for HD anticipation. With the recent discovery of HD cells in the deep layers of the entorhinal cortex (EC) adding to the already sizeable number of HD areas, I will then discuss some effects of propagating (population-coded) HD signals through multiple layers of simulated neurons. Filtering due to synaptic and membrane dynamics appears to be sufficient to account for the observed ~25ms latency between representations in successive HD areas. This same process suggests that SRA in the vestibular inputs can not only explain anticipation in HD cells but could improve tracking performance in distant layers like EC. A corpus of >14 hours of rat tracking data is used to provide an unbiased assessment of model performance.

13/12/05 How to build your own turbocharged linear associative network

David Sterratt

In this talk I'll explain how to assemble your own linear associative network using basic components available from all good Higher grade and equivalent maths courses. I'll then discuss how use a signal to noise analysis to benchmark the performance of your network. This analysis shows how you can tweak the learning rule and threshold to give a truly turbocharged linear associative network that will be the envy of your friends.

I'll discuss whether the brain might employ any of these techniques, and if I've time, I'll mention an idea that links with Mark's talk last week about recognition nets.

13/12/05 Findings From a Striatal Network Model

Martin Guthrie

In my previous ANC talk I presented the construction and setup of a small network of biophysically realistic striatal principle neurons. This network was used to demonstrate learning of navigation of a small grid.

In this talk I will show how the connectivity of the network was improved to allow navigation of larger grids and present findings of how variation of the neuronal parameters affects learning in the network.

05/12/05 Storage capacity of networks for familiarity discrimination

Mark van Rossum

It has been argued that familiarity and episodic memory are processed and stored in different systems. Bogacz has proposed a network + readout to store familiarity. Here we explore the capacity of this network when the stored patterns are spare, consistent with the known coding in the brain.

Work together with Andrea Greve, David Willshaw and David Sterrat.

05/12/05 Data Capture and Classification software for the Brainwave Project

Stephen Henderson

The aim of the Braiwave project is to develop a novel neural assay for compound screening. It is based on transgenic Drosophila which express the calcium-sensitive luminescent protein Aequorin in a specific region of the CNS (the mushroom bodies). When trace readings are taken of the rate of photons being emitted over time they show a regular pattern of oscillations. If the fly brain is exposed to certain chemical compounds this pattern changes to a new stable state with different amplitude and period. The ultimate goal of the project is to develop a software classifier to recognise these patterns and thus identify which compounds are present in a given trace recording.

I will discuss the work I have been doing over the past year in developing the data-capture software for recording and storing the photoluminescense traces and how we have got around the problem of cross-platform compatibility. I'll also go over some potential techniques for extracting the important spectral features of the oscillation patterns for the classification software.


15/11/05 Approximate Information Maximization: a Summary

Felix Agakov

In this talk I will briefly review the most practical aspects of my PhD work on variational information maximization in noisy channels.

Specifically, I will briefly summarize at least five common approaches to maximizing approximations of the mutual information in noisy channels, and show how they may be generalized or improved by using our approach. Then I will show how a choice of richer variational distributions helps to produce tighter lower bounds on the mutual information. Finally, I will review advantages and disadvantages of using our method for practical tasks, including clustering, binary stochastic coding, data visualization, and error correction.

08/11/05 Drosophila Behaviour at the ANC

Dean Baker

I will be giving an outline of the Drosophila based research I've been doing at the ANC for the past two years. This work has focused primarily on the gravity response of flies but depends on the integration of molecular and bioinformatics techniques. After describing the original genetic screen, I will go on to discuss our attempts at mapping neuropil with gross behavioural defects. Several new avenues of work have been identified during this process and I'll describe some of the video tracking paradigms in development and the potential for future imaging projects of the Drosophila brain.

11/10/05 Model, models and models in computational neuroscience

David Willshaw

I thought that at the start of the new year and with the arrival of new people, I should give a summary talk of my research work. Accordingly, this is a brief gallop through my career, featuring amongst other things, the Associative ('Willshaw' Net), the Elastic Net, a few senior citizen-type remarks, an old photo and a short clip of how axonal guidance really takes place. This is a very condensed version of my talk to the Summer School earlier this year; my apologies to those of you who have already heard it.

04/10/05 Orientation-Contingent Color Aftereffects in A Self-Organizing Model of Visual Cortex

Jim Bednar

James A. Bednar and Julien Ciroux

The McCollough effect is a surprising visual phenomenon that reveals interactions betwen orientation and color processing in the visual cortex. After staring at alternating red horizontal and green vertical gratings, subsequent monochrome horizontal gratings appear green, while vertical gratings appear red. There are two current theories of the effect: adaptation of neurons jointly tuned to both color and orientation, or increasing inhibition between neurons tuned individually for color and orientation. Neither theory has definitive support, and whether the effect occurs in V1 or in higher areas also remains controversial. In this work, we show that the LISSOM computational model of the development of color blobs in V1 exhibits color aftereffects that are strikingly similar to those in humans, as a result of the same Hebbian learning processes that drive long-term development. In the model, both jointly and singly tuned neurons contribute, but the singly tuned neurons have a much stronger effect because they are much more numerous. The model thus suggests that both of the main theories are correct, but that jointly tuned neurons are not necessary for the effect. The results further suggest that the effect is an unavoidable by-product of an ongoing process of decorrelation, via Hebbian learning of the statistics of natural scenes at a wide range of time scales.

04/10/05 Clustering and Kernels

Amos Storkey

I will reintroduce a generative model for clusters that can be used to partially justify spectral clustering procedures, and suggest different clustering kernels. I will also look at how such models can be used to build meaningful generative kernels for Gaussian process methods.

27/09/05 Approximate Methods for Gaussian Process Regression: A Survey and an Empirical Comparison

Chris Williams

Basic Gaussian process regression scales as O(n3), where n is the size of the dataset. In this talk I will describe a number of approximation methods, and show the results of some empirical comparisons

27th September 2005 05

Lars Feuerbach

20/09/05 Measuring calcium in Drosophila brains

Jane Ewins

Transgenic Drosophila expresses the calcium-sensitive luminescent protein Aequorin in specific regions of the CNS. The fly brain produces an oscillating signal from the from activity of the gated calcium ion channels of neurons which is detected by a calcium sensitive luminescent protein . The assay requires the dissection of the fly brain, which is then bathed in a solution and is ready for treatment of test chemicals. Changes in inter-neuron calcium ion concentrations have distinctive and predictable oscillating patterns that can be altered by treating the brain with specific compounds that affect neurotransmitter pathways. The talk will also include a real-time demonstration of the collection of calcium signalling data from the brains of Drosophila.

20/09/05 Binocularity in reading isolated words and text

Richard Shillcock

Much research into the reading of isolated words and text has assumed a cyclopean viewpoint, ignoring the fact that we normally read with both eyes. This situation is starting to change. I will review some of the data that currently make binocularity a programmatic issue in reading research, and offer a new interpretation of some recently reported developmental data in binocular reading.

13/09/05 More about a regularised, discriminative approach to predicting protein interactions.

Wolfgang Lehrach

Short well defined domains known as Peptide Recognition Modules (PRMs) regulate many important protein-protein interactions involved in the formation of macromolecular complexes and biochemical pathways. I examine various probabilistic methods to model well what is going on.


Ruth Durie

30/08/05 RIDGES - a suggested Machine Learning approach

Lena Hansson

In this talk I will give a short background to the current problem (to determine if there are any spatial 'factors', 'RIDGES', influencing the regulation of a specific gene, or not), then I will suggest that it might be possible to solve the problem via a Markov Random Field model.

The talk will be more along the lines of a discussion, where, after the initial monologue, I hope to spark a discussion that could help me decide if this model is appropriate, or if there is something better out there.

I will also quickly mention some problems with the training, test and validation data.

30/08/05 Local Linear Regression and Products-of-Sigmoids to model discontinuities

Marc Toussaint

I'll first talk briefly about standard Local Linear Regression, where Gaussian responsibility regions are associated to each local model, as it is used for online learning of inverse dynamics (motor control). Then I'll present a slightly new approach that Sethu and I presented at ICML 05: Instead of Gaussian responsibility regions we use Products-of-Sigmoids to represent complex shaped and sharply bounded responsibility regions. In that way we can model discontinuities which naturally arise in sensorimotor space, e.g. when interacting with objects.

16/08/05 Fast Learning of Sprites using Invariant Features

Moray Allan

A popular framework for the interpretation of image sequences is the layers or sprite model of e.g. Wang and Adelson (1994), Irani et al.(1994). Jojic and Frey (2001) provide a generative probabilistic modelframework for this task, but their algorithm is slow as it needs to search over discretized transformations (e.g. translations, or affines) for each layer. I will describe how we can use invariant features and cluster their motions to reduce or eliminate the search and thus learn the sprites much faster. (Joint work with Michalis Titsias and Chris Williams).

16/08/05 Sex, Flies and no Videotape.

Douglas Armstrong

J.D. Armstrong, D.A. Baker, J.A. Heward and T.C. Lukins

The characterization and quantification of animal behavior is one of the most resource intensive activities in modern biological science. One of the best known examples is the courtship ritual of the small fly Drosophila melanogaster which has been discussed in the academic literature for almost 100 years. It is an exemplary behavior for the functional dissection of the nervous system as it is a natural behavior that requires the animal to integrate and learn from multiple sensory modalities. Object tracking software is extremely good at tracking single moving objects and describing their behavior particularly where previous examples have been well characterized. However, tracking multiple, interacting objects is more challenging. Here we describe the application of a bounding box segmentation algorithm that works well with Drosophila courtship behavior. We show that once segmented, just a few simple parameters are sufficient to classify the behavior in real-time, requiring no off-line storage or post processing.

09/08/05 Multi-compartmental modelling; the good, the bad and the ugly

Andrew Gillies

This will be a dry run of some ideas I will be talking about at the Edinburgh neuroinformatics summer school in two weeks. The talk is a short "HOWTO" do compartmental modelling and follows a case study in the basal ganglia. I only expect to cover passive properties and some initial issues in adding active properties in this talk, so I apologise if it feels like it terminates prematurely. Feedback would be most welcome.

02/08/05 MEDPUB - Maximum Entropy Decisions about PUBlications

David Strerratt

In a departure from my normal line of work, I'll talk about a system developed during an undergraduate honours project to collect new journal articles from PUBMED and rank them according to the user's preferences. A number of algorithms can be employed to achieve this, including Naive Bayes and Maximum Entropy. I'll compare the performance of the two; Maximum Entropy is considerably better. Along the way, I'll describe the Maximum Entropy algorithm for the uninitiated.

2nd August 05

Michael Schouten

I will discuss three independent studies (published in Science on 15 April 2005) that claim to have identified a functional polymorphism implicated in Age Related Macular Degeneration (ARMD). I will review the statistical methodology and biological assumptions used in the studies and explain what implications these results have for future fine-mapping studies.

26th July 05

Andrew Pocklington

26/07/05 Clustering: Spectral Approaches and Recent Advances

Felix Agakov

In the first part of the talk I will briefly speak about a motivation for spectral approaches to clustering. Then without going too deeply into technical details, I will summarize recent work on determining connections between spectral clustering and kernelized k-means. Finally, I will quickly review recent work on determining the number of clusters in the k-means, and outline some of the ongoing work addressing this problem from the spectral and generative perspectives.

My talk will be partially based on the selected highlights of the recent PASCAL workshop on Statistics and Optimization of Clustering.

19/07/05 "There's many a slip...."

David Willshaw

On use to which computational models of biology have been put is to suggest alternative ways of accounting for certain biological facts in terms of specific mechanisms. By doing this, the 'It must be true' line of argument can be defused.

I will describe some work that I have just started where, to my mind, the claimed link between the data and the usual explanation is fairly tenuous.

Much evidence has accumulated that nerve connections are made on the basis of interactions between particular types of molecules carried by the participating cells. Based on some tissue culture work (which I shall summarise), it has long been held that the interactions between the EphA receptors and the ephrinA ligands, which are thought to determine the ordering of connections from axons along one particular axis of the retina onto its targets, are repulsive in nature.

This is an important conclusion which, if true, will constrain our thinking about possible ways by which nerve cells can develop ordered patterns of nerve connections. I shall show that the idea of repulsion does not follow directly from the evidence and I describe a simple computational model provides an alternative explanation.

I shall then go on to describe some recent experiments that promise, as always, to make the problem more complicated than it first seemed.


Walter et al Development Vol 101, 685-696 (1987)

Hansen et al Neuron, Vol 42, 717-730 (2004)

19th July 05

Wolfgang Lehrach

12/07/05 Latency and cortical computation

Mark van Rossum

Cortical computation can be very fast (in the case of visual processing, one needs 150 ms to categorize an image). But in less optimal conditions, such as under low contrast, latencies and processing time can be much longer. What can explain this extra computation time? Simple networks can not explain these effects. We continue our exploration of recurrent circuits that can cause long latencies and research their computational role.

12/07/05 Density Traversal Clustering

Amos Storkey

Spectral Clustering has proven to be a powerful clustering tool, utilising the overall distribution of relative distances to help locate sample points which belong together. However spectral clustering suffers from a number of problems. First it is an approach which is not based on any underlying density model: all of its calculations are based on the sample points and ignore the underlying space within which the data sits. There is no generative framework for the data. Second there are border effects that are undesirable, while at the same time, some overlapping clusters are hard to separate.

In this talk I will introduce spectral clustering, review the relationship it has to Markov relaxation on a graph connecting the sample points. I will then suggest Density Traversal Clustering, a probabilistic model for clustered data using a Markov chain in continuous or other spaces which has the true density as its equilibrium model. Short-time iterations of this Markov chain produce mixing within cluster regions, and the clusters can be learnt using variational Bayesian methods, by making a sample approximation to the integrals involved in computing the Markov transitions. It turns out that this method is similar to spectral clustering except that it differs in the choice of affinity matrix by a Metropolis-Hastings term. The usual form of spectral clustering approximates this approach in the large sample smooth distribution limit.

This is work done jointly with Tom Griffiths.

05/07/05 Feasibility of extracting connectivity info from electron microscopy stacks

Fred Howel

One of the difficulties with modeling the behaviour of neural circuits is that we only have a very rough idea of the detailed wiring diagram. It is possible (but tedious) to gather large amounts of volumetric electron microscopy data from which one could in theory determine the precise shape and position of all synapses, dendrites and axons in a block of tissue. In this talk I'll present some work I've been doing on automated reconstruction and growing synthetic stacks of images for testing the algorithms.

05/07/05 Standards compliants middlewares to empower better research

Marielle Lange

In this talk I will present an integrated environment for lexical resources (resources that concern words and their characteristics), involving both web-based material and stand-alone applications. The first part of the talk will spell out what has already been realized. It corresponds to some minimal solution, achieved with limited means, to establish an open access repository of lexical resources (electronic databases, tables with lexical statistics, programs to compute various kinds of lexical statistics) frequently used for the selection and control of materials in psycholinguistic research. The second part of the talk will outline additional actions that could be taken to facilitate, support and co-ordinate the efforts of the community of researchers (psycholinguists, linguists) involved in the creation and use of lexical resources.


20/06/05 Reading and gender

Richard Shillcock

We will present some predictions concerning the extent to which the divided visual field, beginning with the precise vertical division of the fovea, might be resolved early in the visual pathway and hence be of no significance for the higher cognitive functioning involved in visual word recognition.

We will then present the results of an experiment in isolated Chinese character recognition which demonstrates that the gender of the reader has a significant effect on the ease with which different types of Chinese character may be recognized. Specifically, characters in which the phonological information is on the left are processed more effectively by females, compared with males. The results are interpretable that the hemifield division of information is very precise (dividing a single character) and that the effects of the division extend sufficiently far into the processing for the gender of the reader to have an effect. Foveal splitting is therefore a major determinant of word reading behaviour.

06/06/05 6th June 05

John Quinn

I'll be talking about the use of Switching Kalman Filters to make inferences from time series. In particular, our neonatal intensive care dataset has noisy physiological measurements with many different dynamics, which are often corrupted by artifact. Using variations on the SKF we can make some inferences about the baby and the monitoring equipment, and try to estimate what the uncorrupted measurements would have been.

06/06/05 Towards a Biophysical Computational Model of Striatal Function

Martin Guthrie

The main types of computational models of the striatum have either used lateral inhibition, which has now been shown to have insufficient influence to justify the models it was used in or are more abstract, such as reinforcement learning models.

One of the main proposed functions of the striatum is action selection and reinforcement learning models have shown how this can be done using the actor-critic architecture. Most of these models have a simplistic and non-biophysical representation of the actor.

I am working towards showing that a simple biophysical representation of the actor is capable of learning action selection. This model can then be used to investigate the role of dopamine in learning in the striatum.

31/05/05 Probabilistic Inference Planning

Marc Toussaint

I will talk on an approach to planning based on probabilistic inference rather than estimating value functions. The basic method can be captured as follows:

You have seen Pip successfully getting from state $A$ to state $B$ in $T$ time steps. You only have a prior on Pip's behavior policy but you know the environment. What is the probability that Pip chose a particular action at time $t=0$?

Discovering such probabilities can be used as a basis for action selection. In this paper we extend the probabilistic inference planning (PIP) approach to a level where it is efficiently applicable, comparable and competitive to planning approaches based on value functions. We remove the need for a fixed time $T$; rather we can choose different priors for $T$ and derive the resulting posterior. The choice of different priors for $T$ enables natural solutions to planning problems like ``reach the goal in maximally $T$ time steps''. Efficiency is achieved by a scheme to prune message passing operations that will not influence the final result.

31/05/05 Modelling hormone release for vasopressin cells

Ruth Durie

Vasopressin cells help regulate blood pressure: they must sustain release of antidiuretic hormone (ADH) into the blood under challenging conditions. It's also know they release vasopressin dendritically, probably to aid load balancing: however, this has never been proved, and experimental evidence variously names va sopressin as inhibitory, excitatory or having no effect upon the cells!

By modelling a network of vasopressin cells, perhaps one can find evidence which determines which proposed vasopressin function can have the desired effect, and indeed, whether these cells could be load balancing at all. A model generating spike trains similar to those of the cells has been developed: to harness this into a network, a way to evaluate it is necessary, so this talk will cover the model of ADH release.

Release occurs in the pituitary: the experimental data shows there is a facilitation effect (lasting 20s) followed by depletion. Given the computational constraint necessary for a model to be networked, and the resolution of the experimental data, a simple two pool model of vesicle release can mimic the data.

17/5/05 Brainwave: an in vivo assay for agrochemical and pharmaceutical discovery

Douglas Armstrong

The brainwave project is an example of a truly multidisciplinary project integrating software engineering, computer system design, transgenic neuroscience, pharmacology and machine learning. It aims to develop a novelassay that can be used to assess the effects of chemicals on the nervous system. We will introduce the background science leading to the project and discuss progress to date focussing on the technical development.

17/5/05 17th May 2005

Lena Hansson

The traditional model of gene expression is that specific enhancer, and inhibitor sequences determine the temporal and spatial expression of a gene. However, there are regions in the genome where this model is either broken or augmented by different model(s). These regions/RIDGEs (Regions of IncreaseD Gene Expression) appear to be regulated at the chromosome level affecting multiple genes in a sequential fashion. We are examining microarray datasets to investigate the structure, organisation and function of these putative RIDGEs.

05/3/05 Slow single cell Oscillations in the Subthalamic nucleus

Andrew Gillies

In my last ANC talk I described the emergence of rhythmic bursting in midbrain cells as a consequence of disorders such as Parkinson's Disease. However, there is a discrepancy between the type of tremor related rhythmic bursting observed in the subthalamus of Parkinson's patients or MPTP primate models (5-7Hz) and induced rhythmic bursting in subthalamic slice and culture experiments (1Hz). It may be that these latter experiments are missing key inputs or network influences. Alternatively, they do not capture multiple bursting modes of single subthalamic neurons. Using a multicompartmental model of a subthalamic cell I show that two bursting modes may exist in the neuron, one with a frequency less than 1Hz and a second at around 6Hz. I will also suggest an in vitro experimental approach to test this idea and reveal the higher frequency mode.

05/3/05 Positional Cloning of Genes

Michael Schouten

Although the entire complement of human genes has been mapped, much remains unknown about their function. An important problem in medical genetics is to "clone" (identify) those genes that are implicated in disease. If the disease has a simple genetic basis (i.e. it can always be attributed to a single mutation) then there is an effective statistical approach that locates the gene on the basis of its position in the genome (rather than function). In this talk, I will quickly summarize the straightforward biological concepts that make such an approach possible. I will focus on how the gene responsible Huntington’s Disease was cloned. I will then discuss the problems that arise when trying to apply the same methods to genes with a complex genetic basis, but will review how Hidden Markov Models can be used to overcome some of the obstacles.

29/03/05 Best forgotten: optimal weight decay in linear associative memories

David Sterratt

Most neural network models of memory can learn but not forget. As more patterns are stored in the network, the performance of the network decreases. The number of patterns that can be stored before performance falls below a certain level is the capacity of the network.

In contrast, networks that exhibit a palimpsest property can store new memories continuously. The recall error of a pattern increases with the time since it was stored. In this case the capacity is the number of new patterns that can be stored before the performance of a particular pattern falls below threshold. Weight decay and weight limiting are two methods of turning Hopfield networks into palimpsest memories.

Hopfield networks are a special case of linear associative memory. I show how adding a weight decay term to any type of learning rule in a linear associative memory can turn the memory into a palimpsest. Depending on the level of performance desired, there is an optimal rate of weight decay that maximises the capacity. At the optimal weight decay rate, the capacity is approximately 1/e times the capacity of a network without weight decay. Depending on the learning rule, the capacity is proportional to either the number of units in the network or the square root of the number of units.

I will contrast these results with a recent paper by Fusi, Drew and Abbott on memory retention (Fusi & al. 2005 "Cascade models of synaptically stored memories" Neuron 45, 599-611.

29/03/05 A Review of Probabilistic Hierarchical Clustering

Chris Williams

29/03/05 Variational Criteria of Informativeness

Felix Agakov

In the first part of the talk I will briefly compare approximate approaches to likelihood, conditional likelihood, and mutual information maximization. Specifically, I will show that the variational conditional likelihood training of *noisy* autoencoders may be viewed as the simplest form of the variational information maximization (IM) in the corresponding stochastic channels.

In the second part of the talk I will briefly discuss more general ways to approximate the mutual information, and demonstrate a few applications of the IM to dimensionality reduction and clustering.

15/03/05 Phosphorylation and the regulation of synaptic function

Andrew Pocklington

Phosphorylation plays an important role in controlling the activity of proteins involved in many synaptic processes. These include (amongst others) receptors/ion channels, kinases, phosphatases, and cytoskeletal/cell adhesion molecules. Recent experimental advances have allowed the large scale mapping of the synaptic phosphoproteome. I will present an initial analysis of data obtained for a set of 25 kinases and 93 potential substrates covering a broad range of synaptic function. Although this is still a relatively small data set, I will try to use features identified by the analysis to sketch the broader picture of synaptic regulation via phosphorylation.

22/02/05 Latency and cortical computation

Mark van Rossum

(with Matthijs van der Meer and Mike Oram)

Cortical computation can be very fast (in the case of visual processing, one needs 150 ms to categorize an image). But in less optimal conditions, such as under low contrast, latencies and processing time can be much longer. What can explain this extra computation time? Simple networks can not explain these effects. We present an adaptive circuit with a signal dependent latency that start to explain many of the observed phenomena. (Provided my own latency does not get in the way.)

15/02/05 Label-matching rather than neighbour-matching underlies the formation of ordered nerve connections.

David Willshaw

It is widely held that the ordered mappings of nerve connections that are found, for example, between retina and optic tectum or superior colliculus in vertebrates are a result of several mechanisms acting in concert. One view is that mechanisms involving cell-to-cell recognition using the molecular labels carried by individual cells set up a very rough initial map, which is refined by means of mechanisms of neighbour-neighbour interactions probably driven by correlated neural activity.

Here I argue that the development of the ordered map is due predominantly to molecular mechanisms: the guiding principle behind the formation of the map is that axons interact amongst themselves according to the molecular labels that they carry; this allows them to re-create within the ordered projection pattern that they form on the target the gradients of molecular labels distributed across the retina.

According to this interpretation, this molecular mechanism is responsible for the flexibility with which neurons reorganise their connections during development and the degree of precision in the final map. Once the map has been set up, activity-based mechanisms act to minimise the overlap between individual retinal projection fields; in the case of innervation of the target by two separate populations of cells, this leads to the formation of ocular dominance stripes.

In this talk I concentrate on my analysis of the series of retinocollicular maps found in genetically modified mice involving the knockin and knockout of certain EphA receptors that are thought to act as one of the retinal labels in this system (Brown et al., Cell, Vol. 102, 77-88, 2000; Reber, Burrola and Lemke, Nature, Vol. 431, 847-853, 2004). These maps are predicted by the marker induction model (Willshaw & von der Malsburg, Phil. Trans. Roy. Soc. B, Vol. 287, 203-243,1979) which I have extended to incorporate recent findings about the possible identity of the markers (labels). Quantitative predictions can be made, particularly about the precise nature of the gradients of ephrinA ligands distributed over the colliculus in this series of experiments and on the possible effects of applying the same knockin and knockout manipulations to the EphB receptors.

15th February 05

Wolfgang Lehrach

Short well defined domains known as PRMs regulate many important protein-protein interactions involved in the formation of macromolecular complexes and biochemical pathways. Since high-throughput experiments like yeast two-hybrid and phage display are expensive and intrinsically noisy, it would be desirable to more specifically target or partially bypass them with complementary in silico approaches. I present a probabilistic discriminative approach to predicting PRM-mediated protein-protein interactions from sequence data.

01/02/05 - tools for integrating 3D neuronal reconstructions

Fred Howell

Having a detailed wiring diagram of brain circuits and substantial numbers of 3D reconstructions of neurons would help immensely with modeling their operation - but surprisingly few of the reconstructions of populations and cells which have been done by neuroanatomy labs are available electronically. And even where they are, the data from different labs and brain regions are hard to compare and combine, with cells floating in free space and not tethered to 3D landmarks.

In this talk I'll describe ongoing work on, a collaboration with neuroanatomy labs in Rutgers, Oslo, Krasnow, Brown and Penn to develop neuroinformatics tools and databases to help improve data collection and sharing of neurolucida reconstructions of cells and populations in different areas of the rat brain.

01/02/05 Flexible density models

Amos Storkey

I will describe some current research in flexible density models including the use of Gaussian process equilibrium density models.

25/01/05 The Optimal Viewing Position (OVP) effect in visual word recognition: theories and a model

Richard Shillcock

A single visually presented word is processed more effectively if it is fixated at a particular point along its length - the Optimal Viewing Position. A large number of studies have addressed this phenomenon and a lot is known about the associated reading behaviours. I will review a number of factors which have been argued to contribute to this effect, and I will present a model of the establishment of orthographic representations which allows several different relevant dimensions to be manipulated and allows the effect to emerge.


Matthijs van der Meer

The firing of rat head-direction (HD) cells can anticipate the animal's actual head direction by up to 80 ms. The degree to which a given HD cell shows such an anticipatory time interval (ATI) is correlated with how much its tuning curve distorts during turns, as well as with tuning curve bimodality when still. This phenomenon has most recently been modelled using offset connections from one HD ring to another; however, this type of explanation conflicts with anatomical data showing that the most anticipatory HD cells appear to be driving less anticipatory ones. I will present some work in progress (with Mark van Rossum) on alternative explanations of anticipation and tuning curve deformation in the HD system which do not require elaborate offset connections and are consistent with anatomical and lesion data. The mechanisms considered include HD cell firing rate adaptation, adaptation in vestibular neurons, and vestibular post-inhibitory firing.


Dean Baker

For the past year and a half, Douglas Armstrong and I have been developing a research program that focuses on the use of genetic dissection & Drosophila melanogaster for investigating cognitive function in animals. I\x{2019}m going to give an overview of what genetic dissection is and the types of behaviour seen in the fruit fly which make it an interesting model to work with. I\x{2019}ll also discuss some of the practicalities for studies of this nature, before describing how neuroanatomy and structure/function relationships of the central brain are currently being investigated.


John Quinn

Physiological monitoring of newborn infants is difficult: false alarms, data overload and artifact corruption are all problems in the intensive care unit. I will talk about the application of probabilistic modelling to address some of these problems - specifically on using extensions of the hidden Markov model to identify different states of health or corrupt data at the cotside. I will show some results from the system in progress, and also some pictures of babies.

11/01/05 Is slowness a learning principle of visual cortex?

Laurenz Wiskott

Slow Feature Analysis (SFA) is an algorithm for extracting slowly varying features from a quickly varying signal. It has been shown in network simulations on 1-dimensional stimuli that visual invariances to shift, scaling, illumination and other transformations can be learned in an unsupervised fashion based on SFA [1].

More recently we have shown that SFA applied to image sequences generated from natural images using a range of spatial transformations results in units that share many properties of complex and hypercomplex cells of early visual areas [2]. We find cells responsive to Gabor stimuli with phase invariance, sharpened or widened orientation or frequency tuning, secondary response lobes, end-stopping, and cells selective for direction of motion.

These results indicate that slowness may be an important principle of self-organization in the visual cortex.

[1] Wiskott, L. and Sejnowski, T.J. (2002). Slow Feature Analysis: Unsupervised Learning of Invariances. Neural Computation, 14(4):715-770.

[2] Berkes, P. and Wiskott, L. (2003). Slow feature analysis yields a rich repertoire of complex-cell properties. Cognitive Sciences EPrint Archive (CogPrints) 2804,

11/01/05 How to master your brainwaves: a computational approach using Independent Component Analysis.

Arnaud Delorme

In brain wave training, typically subjects attempt to control brain activity at a specific frequency (for instance they attempt to decrease 3-8 Hz oscillations) at one passive recording electrode placed over subjects\x{2019} scalp (for instance, near the top of the head). The goal of brain wave training can be 1) to control a cursor for paralyzed patients that cannot even move their eyelid to communicate with the external world or 2) to train the brain to produce rhythms away from a clinicopathologic EEG signature for instance to attempt to cure epileptic crisis. However all known experimental procedures rely on training subjects to produce specific rhythms at a single scalp electrode, and little is known of the underlying brain dynamics producing such changes. My research project is to use more recording electrodes (up to 256 simultaneously) and have subjects train brainwaves in one specific brain area and not at a single scalp electrode. Based on my previous research that we pioneered at the University of San Diego California, I will outline how this is possible using a brain wave source separation method called independent component analysis (ICA). Using multiple data channels to compute the brain source activity in real time with ICA produces a clearer signal. As I have already observed in one pilot subject, this improvement is likely to allow subjects to better control their brainwaves and will help us to better understand brain dynamics at the origin of pathological mental states.

14/12/04 Determining optimal pedigree structures for genetic association studies

Michael Schouten

In this talk, I will provide an overview of my current research, which entails developing a likelihood-based model to estimate good pedigree structures for use in fine-mapping a complex trait. I begin by discussing marker-based mapping of complex traits in the broader context of "functional genomics", and in particular highlight the emerging synergies between microarray data and SNP markers for genetic mapping. I then discuss the benefits of a likelihood-based approach using graphical models for learning pedigree structure, emphasizing the potential importance of the Fisher Information metric in evaluating structure.

14/12/04 RIDGE (Regions of Increased Gene Expression) hypothesis

Lena Hansson

This talk will be presenting the latest results regarding the RIDGE (Regions of Increased Gene Expression) hypothesis. This will include an short introduction of the dataset (talking about the sort of data, for instance MHC, that we are expecting to see, and that therefore can be used as a positive control), a repetition or the definition of a RIDGE and the tool built to find them.You will see that although the work is still not conclusive , it actually does finds these positive controls. This would imply that the hypothesis (and the framework built to test is) is correct, although more data is needed.

07/12/04 Fast Learning of Sprites

Chris Williams

Jojic and Frey (2001) provide a generative probabilistic model framework for the interpretation of image sequences using layers/sprites.

In this talk I discuss two methods for speeding up the learning of sprite models. The first method (presented at GMBV'04) uses approximate tracking of the objects in the scene. First, the moving background is learned, and moving objects are found at later stages. The algorithm recursively updates an appearance model of the tracked object so that possible occlusion of the object is taken into account which makes tracking stable. We apply this method to learn multiple objects in image sequences as well as articulated parts of the human body.

For the second method we show that by using invariant features (e.g. Lowe's SIFT features) and clustering their motions we can eliminate the search and thus learn the sprites quickly.

Joint work with Michalis Titsias and Moray Allan.

07/12/04 Disfluency affects expectancy of listeners: Evidence from Event-related potentials (ERPs)

Lucy MacGregor

Conversational speech is full of disfluency (e.g. 'er'; 'um'). How do listeners deal with such interruptions? The current study explores the effect of disfluency on language processing, using Event-Related Potentials (ERPs) as an on-line measure of the comprehension process. Existing behavioural research shows listeners to be sensitive to disfluency (e.g. Bailey & Ferreira, 2003; Brennan & Williams, 1995; Fox Tree, 2001). To date, however, the mechanism underlying this sensitivity is unclear. One suggestion is that disfluency signals to the listener that upcoming information is difficult to access. But can listeners use disfluency in a linguistically predictive way?

In the present study we recorded ERPs whilst participants listened to conversational-type sentences, which ended in a highly expected or less expected target word (as determined by a pre-test). Utterances were either fluent or disfluent. Specifically, disfluent utterances had an 'er' preceding the target.

Fluent utterances elicited a classic N400 component (relative increase in negativity peaking approximately 400 ms after the onset of unexpected compared to expected words, cf. Kutas & Hillyard, 1984). However, for disfluent utterances, the amplitude of the N400 was reduced. The results suggest that disfluency reduces the influence of previous context in creating expectancies for subsequent linguistic information. I argue that disfluency is used in linguistic prediction, signalling to the listener not to rely on expectations formed from prior context.

23/11/04 Self-organization of color opponent receptive fields and laterally connected orientation maps in V1

Jim Bednar

Joint work with Judah De Paula at the University of Texas at Austin

Although much is known about orientation processing in primate visual cortex, relatively little data is available regarding color selectivity. Recent experiments using optical imaging add support for the idea that color-selective cells in V1 are grouped into spatially separated "blobs", with the remaining cells having largely monochromatic preferences, e.g. for orientation. It is not yet known, however, whether neurons in these blobs make the long-range patchy lateral connections typical of monochromatic orientation-selective cells. Using a self-organizing model of V1 with natural color image input, we show that realistic color-selective receptive fields, color blobs, and orientation maps develop. Connections between orientation-selective cells are orientation specific, matching previous experimental results, and the model predicts that color-selective cells will primarily connect to other cells with similar chromatic preferences. These findings suggest that a single Hebbian learning system may underlie the development of orientation selectivity, color selectivity, and lateral connectivity, and provide concrete predictions to be tested in experimental studies.

23/11/04 Harmonising chorales by probabilistic inference

Moray Allan

I will describe how we used a data set of chorale harmonisations composed by Johann Sebastian Bach to train Hidden Markov Models. Using a probabilistic framework we created a harmonisation system which learns from examples, and which can compose new harmonisations. I will explain how we made a quantitative comparison of our system's harmonisation performance against simpler models, and play some harmonisations generated by the system.

09/11/04 Rhythms in the Deep

Andrew Gillies

Rhythmic oscillations that are observed in neurons deep in the midbrain are generally associated with certain disorders of the nervous system (for example, Parkinson's Disease). Treatments of the disorders often target these areas, either surgically, chemically, or electrically. What is the origin of the rhythms? We have previously shown how the network arrangements in key areas of the basal ganglia can dispose the system to rhythmic oscillations under Parkinsonian conditions (Gillies & Willshaw 2002). Here I will invistigate whether single cells can also play a role in the generation of abnormal rhythms.

09/11/04 Unsupervised Learning: Probabilistic and Information-Theoretic Perspective

Felix Agakov

First I will review basic approaches to unsupervised learning related to maximization of the likelihood in generative models and maximization of the mutual information in noisy channels. Then I will discuss similarities and differences between the approaches in the context of induced optimization problems, and demonstrate the performance of both methods for a simple clustering task. If time permits, I will then speak about a few simple observations regarding optimization of Becker and Hinton's Imax criterion, and comment on Becker's empirical observations related to Imax learning.

26/10/04 The effect of differentially-attenuated inputs in linear associative memories

David Sterratt

Synaptic inputs tend suffer different different degrees of attenuation depending on where they lie on the dendritic tree. I consider the effect of differentially-attenuated inputs on the performance of a linear associative network with real-valued weights. The signal to noise ratio (SNR) between units that should be "high" and "low" for any given pattern is a measure of the performance of the network. It has been calculated for arbitrary learning rules by Willshaw and Dayan. I extend their expression for the SNR to account for arbitrary attenuation of each input, or random attenuations drawn from a known distribution. The new SNR accurately predicts simulation results.

As an approximation to real neurons, I study the case where there is a uniform distribution of inputs along a dendritic tree and the attenuation depends linearly on distance. I find that when the distal inputs are attenuated twice as much as the more proximal ones, the SNR of the network is decreased by a factor of less than 4% and the maximum reduction in performance (for an infinitely steep gradient) is 25%. This is a curious given that at least some neurons seem to scale their inputs to compensate for distance. Other arguably more realistic distributions of inputs also suggest the performance hit due to differential attenuation is surprisingly low.

19/10/04 Identifying functional components in biological networks

Andrew Pocklington

This will essentially be an extension of my previous talk in which I presented ways of predicting molecular involvement in biological function/disease. I will describe a set of methods based on the analysis of molecular interaction networks and show how their performance may be simpy evaluated. Applied to a post-synaptic protein complex, these methods are used to extend functional, behavioural, and disease-related protein sets. Evaluation shows some of these predictions to be surprisingly robust, and clearly highlights others as far less reliable.

12/10/04 Fovea is split or not? An examination of Chinese semantic radical combinability effects with lateralized cues

Janet Hsiao

In Chinese character structures, a semantic radical, usually on the left, bears information about the meaning of the character. Recent evidence has found a facilitation effect of radicals with large combinability in a character categorization task (Chen & Weekes, 2004). Here we report our attempt to utilize this effect to examine the split fovea theory, which proposes a precise midline splitting in the human foveal representation and the consequent contralateral projection to the two hemispheres, by presenting lateralized cues during a character semantic transparency judgment task. We showed that when semantic radical position was controlled to be on the left of the characters, there was a significant interaction between cue position (left vs. right) and radical combinability; a left, but not a right cue, facilitated the recognition of opaque characters and hence eliminated the combinability effect among opaque characters. The results support the split-fovea theory, where a left cue attracts attention to the semantic radical to the left of fixation and hence facilitates semantic judgment, while a right cue does not.

05/10/04 What is the size of the unit we use for print-to-sound conversion: graphemic or susbyllabic?

Marielle Lange

Pring (1981) constructed an experiment in which she alternated the case of the letters. In one condition, she left the units functional for pronunciation intact (CHurCH) and in another she disrupted them (ChuRCh). She then observed the impact of the two kinds of case disruption in a lexical decision task for words, pseudohomophones (strings as "cherch" which are not words in English but which have the same pronunciation as an existing word, here "church") and nonwords (strings as "chorch" which are not words in English, neither in their spelling nor pronunciation). The discovery that disruption inside a unit functional for pronunciation had an impact on readers' performance led her to conclude that these units are represented in the reading system. A problem with this proposition, however, is that the experiment did not provide unambiguous evidence for the grapheme. The intact or disrupted unit was at the same time the grapheme and the subsyllabic unit (initial consonants, vowel, final consonants). In this talk, I will report an experiment that does not include that confound.

05/10/04 Learning representations: The genome as a self-organized representation to model phenotypic variability --- and how that relates to things like ICA

Marc Toussaint

I will first briefly introduce to some of my work on the evolution of genetic representations. The key point is that when there exist multiple genetic representations of the same phenotype, then there is an implicit selection pressure between them which is related to the Kullback-Leibler divergence between their phenotypic variability and the fitness distribution---genetic representations evolve according to how well they allow to model a desired phenotypic variability.

I will then discuss a broader point of view by considering a general representation problem. The choice of genetic representation and similar problems can be viewed as special cases. I'll argue for incremental approaches to learning representations. When the data are strings, grammatical representations can be incrementally build up. But what if the data is continuous?

28/09/04 Charge based fluctuation analysis

Mark van Rossum

We present a new way to derive synaptic properties from fluctuations in the synaptic currents.

In particular when the synapses are remote and the currents are filtered by the cable properties of the neuron this new approach improves older methods.

The method also indicates that fluctuation in the charge might be substantial and effects the reliability of synaptic integration in a way which was not considered so before.

28/09/04 Towards a biophysical model of reinforcement learning

Martin Guthrie

Since the realisation that the dopamine signal in the striatum is very similar to the reward prediction error signal in reinforcement learning models there has been much interest in modelling learning in the basal ganglia as a reinforcement learning system. There have, however, been some problems with relating the theoretical requirements of reinforcement learning to the known biophysical parameters. In this talk I will highlight some of these problems and attempt to show how a more realistic computational model of learning in the striatum could be constructed.

21/09/04 Nature/Nurture in the formation of ordered nerve connections?
Examination of some actual evidence for the different possible mechanisms dreamt up by modellers

David Willshaw

In proposing a computational model as a solution to a biological problem, it is essential to examine the supporting experimental evidence. In this talk, I review the problem under discussion, which concerns how the nervous system wires itself up, together with one possible solution, the inductive hypothesis (discussed in earlier talks). I then examine some old and some new experimental results. The aim is to show how these results support/refute the inductive hypothesis or other competing hypotheses.

21/09/04 Detecting recombination and rate variation in DNA sequence alignment.

Mark Chiang

Hidden Markov models have become one of the dominant tools in biological sequence alignment. In this talk, I will present the methodology and the results of detecting recombination and rate variation in DNA sequence alignment.

14/09/04 Title: Is The Grid useful for neuroinformatics?

Fred Howell

"The Grid" started out as a way for CERN to distribute their petabyte data analysis requirements from the large hadron collider amongst compute facilities in other reseach centres, and was then promoted as a tool for new ways of doing e-science in other disciplines - bigger, better and more ambitious.

I'll outline some of the projects from the recent EPSRC "All Hands" conference ( and discuss whether / when "griddy" ideas might be useful for computational neuroscience.

14/09/04 Patterns of Linkage Disequilibrium (LD) in the Human Genome

Michael Shouten

The presence of LD (i.e. correlations between neighboring alleles on a chromosome) has strong implications for both disease mapping and understanding evolutionary history.In this talk, I will provide a high level introduction to modeling LD in these two contexts. I will also summarize the results published in several key LD-based studies of the human genome.

07/09/04 Emergent effects in visual word recognition

Richard Shillcock

I will talk about a number of behaviours that have been observed in the reading of single words and texts. In the literature these effects have been variously ascribed to perceptual learning or to processing "strategies" of indeterminate status. I will show that these behaviours robustly emerge from a connectionist model that has learned a shift-invariant mapping from orthographic features to letters over a realistically large lexicon of monosyllabic words. Any processing strategies are in this sense optimal. I will also show that certain dyslexic behaviours emerge from this model as a result of non-specific impairment.

07/09/04 Models of the rodent head-direction cell system

Matthijs van der Meer

Head-direction (HD) cells, which fire selectively and persistently when the animal is facing some direction in the horizontal plane, have attracted interest from systems-level modellers as a relatively simple case of stable attractor dynamics in the brain. A continuous ring attractor network, with a Gaussian-shaped activity packet stable at any place in the ring, can account for the most salient experimental observations. However, recent (as well as some not-so-recent) experiments have revealed a number of odd effects which show that such models are at best only a small part of the story, and at worst in fact wrong. I will be reviewing the features of the most recent HD cell models and the ways in which they can and cannot account for experimental data, along with some demonstrations of MATLAB implementations. This will serve as an introduction to a quick discussion of some of my own electrophysiology experiments currently in progress.

18/08/04 Modelling Burst Firing of Vasopressin Cells

Ruth Durie

Vasopressin cells display a distinctive pattern of high frequency burst firing. Firing is very all or none: isolated spikes are rare. I will present a new model mimicking firing patterns in these cells.

18/08/04 Great Moments in Unsupervised Learning

Chris Williams

I will discuss some models/algorithms for unsupervised learning including manifold learning, Boltzmann machines, Infomax (Linsker), IMAX (Becker and Hinton), and ICA. Come along to find out which are considered to be "great moments".

10/08/04 Diversity among Subthalamic Nucleus Projection Neurons

Andrew Gillies

Neurons of the rat subthalamic nucleus, the key excitatory nucleus in the basal ganglia, appear anatomically as a single population of projection neurons. They also exhibit a common set of passive electrical properties. Yet this population shows significant diversity under a wide range of physiological conditions. Some neurons exhibit slow depolarising responses upon stimulation, others slow action potentials (Nakanishi et al 1987). Neurons may exhibit short post hyperpolarisation responses or long responses (Hallworth et al 2003). Some neurons exhibit slow rhythmic bursting, whereas others do not (Beurrier et al 1999). What are the key cell properties that underlie this diversity? I will present a model of the rat subthalamic projection neuron that exhibits many of these characteristic behaviours, and demonstrate which parameters are the key candidates in this diversity. I will focus on the modelling techniques used to capture the subthalamic neuron characteristics and to explore physiological diversity.

10/08/04 Structural Equation Models for Bayesians

Amos Storkey

Structural equation models are used in many domains, from Economics to models of connectivity in the Brain and in fMRI. In this talk I will provide an introduction to structural equation models. The focus will not be on _how_ to use them, but what they are, their inherent properties and why to use them. The theory is that once those are described properly, then how to use them is an exercise in Bayesian methods.

Structural equation models (SEM) are usually described as a set of linear simultaneous equations with a variety of 'statistical' methods available for determining the values of the parameters given data.

In this talk I take a more generative approach. A generalisation of structural equation models can be defined which provides a third type of graphical model. These are different from belief networks (Bayesian networks) and Markov networks (undirected graphical models). They have different characteristics and different independence relationships. They have belief networks and Markov Networks as special cases.

I will provide a description of the generative process for the generalised SEM, and discuss the reasons why it might be useful. I will also describe the linear Gaussian SEM in these terms but then give an alternative interpretation as the long term average activity of an equilibrium Bayesian network. This is an appealing interpretation in an fMRI setting for example. Some comments regarding the issues in SEM modelling (inference and learning) will be covered.

03/08/04 Fast Unsupervised Greedy Learning of Multiple Objects and Parts from Video

Michalis Titsias

I will talk about how our earlier work on learning multiple objects in images can be greatly speeded up for video sequence data by carrying out approximate tracking of the multiple objects in the scene. Our method is applied to raw image sequence data and extracts the objects one at a time. First, the moving background is learned, and moving objects are found at later stages. The algorithm recursively updates an appearance model of the tracked object so that possible occlusion of the object is taken into account which makes tracking stable. We apply this method to learn multiple objects in image sequences as well as articulated parts of the human body.

03/08/04 Does meaning change your memory?

Andrea Greve

Throughout our lives we learn general knowledge about the world (semantic memory), whilst also experiencing specific events that occur at a particular time and place (episodic memory) (Tulving, 1972). Semantic and episodic memory are regarded as distinct memory systems. However, it is likely that these systems interact, so that one system may benefit or interfere with the other. Event-related potentials (ERPs) permit the investigation of the neuralcorrelates of semantic and episodic memory. Event-related potentials provide an online and direct measure of the brain?s neural activity: they are an excellent tool to investigate the time course of cognitive events, and also provide some indication of brain regions in which these events occur. I will present an ERP study that investigates how semantically related and unrelated stimulus material alters episodic memory performance. Of interest is whether changes in memory performance are supported primarily by recollection or familiarity processing. Therefore, I will discuss ERP components that index familiarity and recollection.

20/07/04 Identifying functional components in biological networks

Andrew Pocklington

To understand a particular biological (dys)function at the molecular level, known biochemical components must be placed in their natural context - the complex web of interactions occuring in vivo. Where information concerning molecular components is incomplete or masked by noise (as it usually is), strategies must be developed to identify those genes/proteins involved. These are expected to form a connected subset within the network of all (local) interacting molecules. I will outline a method which uses network structure to direct the search for functionally relevant molecules.

20/07/04 Information Maximization in Intractable Channels

Felix Agakov

The key idea of information maximization is to choose a mapping from source variables to response variables such that the output contains as much information about the transmitted inputs as possible. Despite the conceptual simplicity of the statement, the problem is in general computationally intractable in all but special cases.

In the first part of the talk I will motivate a simple lower bound on the mutual information for sigmoidal networks and compare it with several techniques suggested for stochastic neural coding. I will also briefly describe an extension of the model to the temporal case.

In the second part of the talk I will generalize the simple bound and show that many popular techniques (such as Kernel PCA, Gaussian Process Latent variable Models, Information Bottleneck, etc.) arise as special cases of variational information maximization for specific channels.

13/07/04 Spine calcium dynamics during spike timing plasticity and Winnow: a program to separate table of contents wheat from chaff

David Sterratt

Calcium signals in spines are believed to play a critical role in synaptic modification, and a recent in vitro study (Nevian & Sakmann 2004, J. Neurosci. 24, 1689-1699) has suggested how spine calcium transients depend on the relative timing of pre- and postsynaptic activity. Unfortunately, the fluorescent dyes used in the study affect the dynamics of the calcium signals. I use a compartmental model and a simple model of calcium buffering (Sabatini & al. 2002, Neuron 33, 439-452) to address the question of what the calcium signals might be under in vivo conditions with no dye. Along the way, I'll be giving a quick introduction to calcium dynamics.

I was also be demonstrating a program called Winnow, which an undergraduate CS4 student (George Chambers) is working on. The idea is to avoid reading through lots of chaff (boring articles) in email tables of contents and instead look at the wheat (interesting ones). The program uses the Naive Bayes text classification algorithm and adapts to the user's preferences. It will be publicly available soon in an alpha version, and comments on it would be welcome.

13/07/04 Informativeness and Optimal Visual Word Fixation: A Computational and Psycholinguistic Investigation.

Alex Mc Cauley

In this talk I will introduce phenomena associated with reading and isolated visual word reognition and describe the key questions raised. I will then outline work I am currently undertaking.

---------- Background ----------

Human readers demonstrate a preferred viewing location (PVL) in normal reading and a comparable optimal viewing position (OVP) for recognition of isolated words. The OVP is the fixation point in a word that gives the fastest, most accurate recognition of that word. The link between, and factors contributing to these effects are still under debate. The key challenge is to explain the asymmetry of these curves; the PVL and OVP fall to the left of word centre in longer words. What we propose to explore here is the possible role of informativeness.

Some models of reading behaviour suggest that identification of written words is made faster by fixating towards more informative parts of the word, where it is best distinguished from other word candidates (e.g.: Clark & O'Regan, 1998; Legge et al. 1997). Evidence of this effect in humans comes from various studies by O'Regan and others (e.g.: O'Regan et al., 1984; O'Regan and Levy-Schoen, 1987). These studies found an influence of information distribution on the OVP curve, although rarely a complete reversal in the curve for end informative words as might be predicted.

Such a role of informativeness is captured in the Split-Fovea Model of Shillcock et al. (2000). This model assumes that the parts of the words at the left and right of the fixation point are projected to the contralateral hemisphere. They define the optimal fixation point to be where both the left and right parts of the word are the most informative (i.e., shared by the smallest number of words in the lexicon). Using a computational algorithm based on these principles, correct predictions for OVPs in both English, a 'left-to-right' language, and Hebrew, a 'right-to-left' language, have been generated (Shillcock et al. 2000; McCauley, AmLaP, 2002).

However, while the simulation results resemble the nature of the OVP curve, no systematic comparison of the degree of skew in the OVP curves has been made between simulation and human data

------------- Current Study -------------

We propose to directly contrast the model's predictions with human data. First we investigate how OVP predictions from the algorithm actually correspond to human OVPs for particular word lengths. Various parameters of the algorithm are changed and compared, including the incorporation of word frequency.

We will then run a 'moving OVP' experiment in English. The algorithm can be used to generate stimuli, processing words and non-words to generate beginning- and end-informative stimuli sets. The experiment will use forced fixation with a lexical decision and word identification task, in order to directly test predictions from the model of the effect of information profile on location of the OVP.

---------- References ----------

Clark, J.J. and O'Regan, J.K. (1998). Word Ambiguity and the Optimal Viewing Position in Reading, Vision Research, 39, 4, 843-857.

Legge G.E., T.S. Klitz & B.S. Tjan. (1997). Mr. Chips: An ideal-observer model of reading. Psychological Review, 104, 524-553

McCauley, A. K. & Shillcock, R. C. (2002). Hebrew and the Hemispheres: A Computational and Psycholinguistic Investigation of Reading Behaviour. Poster presented at AMLaP-2002, Tenerife, Spain. September 19-21, 2002.

O'Regan, J. K., Levy-Schoen, A. , Pynte, J., & Brugaillere, B. (1984). Convenient fixation location within isolated words of different length and structure. Journal of Experimental Psychology: Human Perception and Performance, 10, 2, 250-257.

O'Regan, J. K., Levy-Schoen, A. (1987). Eye movement strategy and tactics in word recognition and reading. In M. Coltheart (Ed.), Attention and performance XII: The psychology of reading (pp. 363-383). Hillsdale, NJ: Erlbaum.

Shillcock, R., Ellison, T.M. & Monaghan, P. (2000). Eye-fixation behaviour, lexical storage and visual word recognition in a split processing model. Psychological Review 107, 824-851

06/07/04 Exploiting partial connectivity in a feedforward associative memory network

Kit Longden

Reducing the connectivity in a feedforward associative memory network reduces the capacity. In his model of the hippocampus, Marr (1971) suggested an algorithm for choosing the output patterns to compensate for the reduction in connectivity. I will present analysis and simulations showing that the algorithm gives a better performance than networks with a higher, specified connectivity that have stored random patterns, for low connectivities and high memory loads. This superior performance is maintained when a more complex threshold strategy is used in the network with random patterns. Finally, I will return to the hippocampus and discuss the implications for information processing in the projection from CA3 to CA1.

06/07/04 Gravitational Responses: Knockout of Brain Structures & Genes in the CNS

Dean Baker

This talk will look at two ways of studying gravitational responses in Drosophila with molecular techniques in which we can knock out subsets of neurons and even specific genes within the fly brain. I will present preliminary results for a trial aimed at knocking out neurons in the central complex and mushroom body of flies, and outline the use of RNA interference in Drosophila for knocking out the expression of candidate genes. Its hoped that this approach will eventually give us an insight into how gravitational stimulus is processed in the CNS.

29/06/04 Saturation in Classifier Spaces

Bob Fisher

When assessing classification results, based on selection of members from a database (eg a face database), one would like to know what is an achievable classification rate, given the noise level, dimensionality of the feature set and number of classes in the database. As best we can tell, no results exist for this question, although many classification rates appear in different papers. This talk presents an empirical formula that links the number of discriminable classes to the error rate, dimensionality of the feature data and the feature noise level.

29/06/04 Genes 2 Cognition - an update

Douglas Armstrong

The postsynaptic terminal receives signals from neurotransmitters and other receptors that activate signaling pathways critical for learning and memory in mammals. Many of these molecules are known to be important in cognitive, pathological and developmental settings. Our own (and other) studies have identified the composition of the postsynaptic proteome (PSP) with at least 700 member proteins. The evolutionary history of this protein complex has been analysed across 11 species. We have reconstructed a model of this proteome and compared the network properties of member proteins with known functional properties. We have shown that the network properties derived from our model correlate with physiological measurements of processes that underlie cognition in mammals. Predictions based on network properties have been fed through to a large-scale human genetics study and the first data from this study is starting to emerge. I will also present some on-going work and immediate plans.

25/05/04 Learning regularities or distributions of probabilities? Answer from an implicit learning task

Marielle Lange

In previous studies (Lange & Content, 1999), I showed that reading performances are influenced by the probability of the print-to-sound associations (or more specifically, grapheme-phoneme associations). This established that, contrary to what was assumed by a very influential model of reading, some information about the uncertainty of the pronunciation of the graphemes must be encoded in the reading systems. One question that remained, however, was whether readers' performance was influenced by the probability of the expected pronunciation or by the uncertainty of it (captured by a measure of entropy). The distinction is that the probability only takes into account one grapheme and its expected pronunciation when the other takes into account all possible pronunciations of a grapheme. To interpret the effect of entropy as evidence for the encoding of the association probability for all possible pronunciations of a grapheme, it would be necessary to show variations in reader's performance with two graphemes which have exactly the same probability for the dominant pronunciation but differ in the distribution of the alternative pronunciations' probabilities. Obviously, the problem would be difficult to address with graphemes in the context of a reading activity because you can scarcely find two graphemes with exactly the same probability for the regular pronunciation and a different number of possible pronunciations. Here, I decided to approach the problem with an implicit learning task in which the probability of occurrence of the next event as well as the number of alternative events that can happen next are manipulated. With this task, I was able to demonstrate variations of performance as a function of the distribution of probability.

25/05/04 Revisiting the Functional Role of the Dentate Gyrus

Paulo de Castro Aguiar

The dentate gyrus is one of the areas of the hippocampal formation. It receives input from the entorhinal cortex and modulator areas, and sends mossy fibres to the CA3, which as long been proposed to act as an auto-association network (Rolls 1996). The mossy fibres contact CA3 through boutons that are unique among cortical synapses due to their size and location (Chicurel and Harris, 1992). It has then been empirically considered that the mossy fiber pathway provides the major source of excitation to the CA3 pyramidal cells. I will show some evidence that opposes this view. Secondly, I will talk about the role of the dentate gyrus as a support network for CA3 auto-association, present some ideas on how the dentate gyrus may provide orthogonalized inputs to CA3 and show my point of view on how sophisticated things can be achieved using un-sophisticated mechanisms.

11/05/04 Detecting recombination in DNA sequence alignments

Dirk Husmeier

The recent advent of multiple-resistant pathogens has led to an increased interest in interspecific recombination as an important, and previously underestimated, source of genetic diversification in bacteria and viruses.A promising method for the detection of interspecific recombination in DNA sequence alignments is the monitoring of changes in the posterior distribution of tree topologies along the alignment, measured with an appropriate probabilistic divergence score. However, as the number of taxa in the alignment increases the posterior distributions become increasingly diffuse. This blurs the probabilistic divergence signals and adversely affects the detection accuracy. The present study investigates how this shortcoming can be redeemed with a pruning and postprocessing clustering scheme. Moreover, an improved method for significance estimation is presented.An application of the proposed scheme to two synthetic and two real-world DNA sequence alignments illustrates the amount of improvement that can be obtained with the proposed approach. The study also includes a comparison with two established recombination detection methods: Recpars, and the DSS method.

11/05/04 Control of the Timing of the First Spike From a Medium Spiny Striatal Neuron

Martin Guthrie

The striatum is the main input nucleus of the basal ganglia, a set of sub-cortical structures involved in the control of behaviours, especially movement. Lack of physiological evidence for models of the striatum proposed over the last ten years has led to a requirement for new ideas about the nature of the calculations that the striatum performs and how these may be implemented in a biophysical model. I propose that one aspect of the computational function of the striatum is that the timing of the first action potential produced in a train is important in the timing of the gating of behaviours and present a computational model to show how the timing of the first action potential could be controlled by one class of interneurons in the striatum.

27/04/04 Learning for Optimization: EDAs, probabilistic modelling or something else?

Marcus Gallagher

Machine Learning methods have seen some application to solving optimization problems. In general, this involves explicitly modelling the data produced by a search algorithm, and using the model (e.g) to increase the speed of the search, to find better solutions, or to gain insight into the problem by examining the model produced. In the field of Metaheuristics (inc. Evolutionary Computation), density estimation techniques and probabilistic graphical models have been used to perform model-based optimization. This work is usually referred to as Estimation of Distribution Algorithms (EDAs). Model-based optimization has also been considered outside the machine learning community (e.g, response surfaces).

In this talk I will provide a brief overview of learning-based optimization algorithms, and attempt to highlight some major challenges in this research domain. In particular, I will discuss the motivation behind EDAs and describe a simple framework for understanding EDAs. The framework is based on minimization of the KL-divergence between the model (probability distribution) and a distribution that depends on the objective function of the optimization problem.

27/04/04 Connectionist modelling of Chinese phonetic compound pronunciation

Janet Hsiao

In this talk, I will describe a connectionist model designed to reflect some of the anatomy of the visual pathways, notably the precise division of the human fovea and its subsequent contralateral projection to the cortex. The model was trained with a realistically large-scale corpus, mapping from Chinese orthography to phonology. Its behaviour was examined with respect to the known effects in regular and irregular pronunciations and a further experiment regarding gender difference in pronunciation

06/03/04 Variational Information Maximization: an Overview

Felix Agakov

Maximization of mutual information in noisy channels is a common, but computationally difficult technique. In the first part of the talk I will outline a simple variational approach to the problem and discuss its possible extensions.

In the second part, I will speak about relation of the method to a variety of common techniques, including likelihood (and conditional likelihood) training in graphical models and non-linear dimensionality reduction.

Finally, I will briefly outline an application to stochastic neural coding.

30/03/04 Is the web about to get back to its roots?

Fred Howell

Tim Berners-Lee's original plan for the web (in 1989) included a number of nifty features which never quite made it into the actual thing - backwards references, links annotated with verbs, and critically it was intended as a read-write medium not a read-only one. Is the recent enthusiasm for wikis, blogs, and microcontent a step towards finally implementing the original prototype properly? Is there scope for improving the concepts of email and the file system at the same time, without getting trapped in a semantic web?

23/03/04 Possible roles of excitability changes for memory formation

Mark van Rossum

In the workshop I will talk about possible roles of excitability changes for memory formation. In rabbit it was shown that exicitability changes in the hippocampus are strongly correlated with delay conditioning learing. One idea is that the excitability change acts as a eligibility trace. I will show a model which reproduces these effects. However, proving its correctness will be difficult... Joint work with Maya Janowitz (Msc Student last year).

Time permitting I would like to talk a bit about symmetry breaking in memory formation. Joint work with Walter Senn.

16/03/04 Molecular-based models for the formation of nerve connections.

David Willshaw

The mechanisms used by the nervous system to wire itself up - to produce in many cases highly ordered patterns of connections - are largely unknown. Contemporary views are that axons are guided to their termination sites in response to electrical signals or by being directed by the labels carried by the participating cells.

In this talk I present my very recent modelling work describing a model for the formation of nerve connections involving the induction of molecular labels (or markers) from the presynaptic array of cells onto the postsynaptic array. This work extends significantly the work carried out in collaboration with Christoph von der Malsburg (vdM & W, PNAS, 75, 5176 - 5178, 1977; W & vdM, Phil Trans Roy Soc B, 287, 203-243 1979).

Features of the new model are:

It is much simpler than the old one;

It is shown to work for 2-dimensional arrays of cells rather than 1-dimensional arrays;

It shows how a map can be developed while the neural structures themselves are developing.

It accounts for the highly abnormal maps produced by genetically based disturbances to the system.

It accounts for the recent work that identifies the Eph/ephrin system of proteins/ligands as the source of the molecular markers.

09/03/04 Formation of retinal ganglion cell mosaics

Stephen Eglen

Retinal ganglion cells can be classified as on-centre or off-centre, depending on their response to light stimulation. I will present alternative hypotheses for the formation of the spatial distribution of on- and off-centre cells. In my last talk, I presented a statistical test and its results on real data sets. These results suggested a spatial dependence between on- and off-centre cells. In this talk, I will describe why that statistical test cannot be used, and present an alternative approach. This new approach suggests that the two populations can develop with only a minor spatial dependency.

02/03/04 Letter and word processing in visual word recognition

Richard Shillcock

In reading words, how visible are individual letters at differing distances from fixation, and how might this influence word recognition? How easy is it (in terms of millisecond timing) to recognize words when they are fixated at different points along their length? What general principles, computational and psychological, are at work?

I will give an introduction to the state of the art in psycholinguistic studies of how the perception \x{2013} and perceptibility \x{2013} of letters and words interacts in reading. I will include some of the interim results of our own modelling work.

17/02/04 Title: 2 papers from NIPS 16 (December 2003)

Chris Williams

(1) Maximum Likelihood Estimation of a Stochastic Integrate-and-Fire Neural Model, J. W. Pillow, L. Paninski, and E. P. Simoncelli. Available via

(2) From algorithmic to subjective randomness. Griffiths, T. L. and Tenenbaum, J. B. Available via

These two papers use probabilistic modelling methods to address issues in computational neuroscience and cognitive modelling.

23/02/03 Bursting Oscillations in the basal ganglia

Andrew Gillies

Repetitive bursts of action potentials separated by periods of quiescent (termed Bursting Oscillations) are commonly found in the basal ganglia under diseased conditions. In particular, in Parkinson's disease, such activity is observed in both the subthalamic nucleus and globus pallidus. The frequency of bursts is consistent with the frequency of tremor, one of the cardinal symptoms of Parkinson's disease

I will present models of the subthalamic-pallidal circuit pitched at different levels to begin to discuss the origin of these Bursting Oscillations. Are the oscillations causing tremor or representing tremor?

27/01/03 Modelling the temperature-dependence of ion channels

David Sterratt

Ion channels are to the brain what the transistor (or valve) is to electronics*. In the first part of this talk I will give a brief tutorial about ion channels and how physiologists measure and characterise them.

The properties of ion channels depend strongly on temperature. In the second part, I will discuss how temperature is incorporated into ion channel models. The "Q10" is a common way of doing this, and I will discuss whether "thermodynamic" models of channels can be expected to do any better than the Q10.

In the third part of the talk, I will present some work done jointly with Andrew Gillies. The aim of this work is to investigate empirically whether thermodynamic models fit data better than traditional models adjusted by the Q10.

* Robert Cannon's quote, I believe.

20/01/03 Unsupervised Learning for Analyzing Microarray Data

Michael Schouten

In this talk, I will provide a high-level synthesis of several key journal articles that employ clustering, projection and model-based methods to analyze gene expression data. The methods covered in this talk will include k-means clustering, Singular Value Decomposition and Independent Component Analysis. Time permitting, I will also review Likelihood-based approaches for learning Bayesian networks that are intended to describe interactions between genes.

13/01/03 Motion segmentation, tracking and learning in image sequences

Michalis Titsias

First I will talk about methods for doing motion segmentation and tracking of multiple "independently" moving patterns (say objects) in image sequences. Then I will explain how these techniques are related to learning multiple objects from video data and particularly how they can be used to considerably speed up a method that learns mupliple objects from image sequences. I will present a fast algorithm that tracks and simultaneously learns the appearances of moving image patterns and I will give some results for learning multiple objects as well as parts of human body data.

16/12/03 Sex, Flies and Videotape

Douglas Armstrong

Apologies for using the most overused title in the repertoire of behaviour geneticists but I have not seen it used in ANC yet. I will briefly describe two studies that have been progressing over the past 18 months are providing some interesting evidence for functional plasticity and compensatory mechanisms in the fly brain. Both studies reveal a surprising degree of functional compensation in response to genetic and chemical damage affecting up to 10% of the neurons in the nervous system. I will describe both sets of experiments and attempt to gather the evidence for compensation in an invertebrate system.

09/12/03 CA1 Place field activity in the hippocampus: two hands, only one mouth.

Kit Longden

CA1 is the main output station of the hippocampus. What kind information is expressed by CA1 place field activity? There are two inputs to CA1 and it has remained a puzzle what the second, cortical input to CA1 contributes. I will present physiological data that investigates how the two pathways interact, before presenting modelling work with integrate-and-fire neurons in which I explore possible schemes. I propose that the main input from CA3 initially creates the place field in CA1 representing context information. This place field is then gradually associated with features in the input through the cortical pathway

18/10/03 Liquid Computation and Liquid State Machines

Paulo de Castro Aguiar

A key challenge for neural modelling is to explain how a continuous stream of multi-modal input from a rapidly changing environment, can be processed by stereotypical recurrent circuits of integrate-and-fire neurons in real-time.I will be presenting recent work (2002) by W. Maass and co-workers and H. Jaeger on a new framework for neural computation that provides an alternative to previous approaches based on attractor neural networks. It is shown that the inherent transient dynamics of the high-dimensional dynamical system formed by a neural circuit may serve as a universal source of information about past stimuli, from which readout neurons can extract particular aspects needed for diverse tasks in real-time. Stable internal states are not required for giving a stable output, since transient internal states can be transformed by readout neurons into stable target outputs due to the high dimensionality of the dynamical system. At the end I will show and explain some practical examples.I would like to emphasise that I will be giving this talk taking into account that there may be people present that are not familiar with these subjects.

18/10/03 Learning representations for probabilistic (world) models

Marc Toussaint

How may neural systems represent and learn a model of the environment and use such models for goal-directed behavior planning? I will first introduce my approach of a connectionist world model, in which perceptual and motor signals are coupled to a growing, self-organizing layer. Eventually, this layer encodes and learns a model of the perceptual implications of motor activity: The lateral connectivity of the layer, modulated by motor signals, induces neural preactivations which encode the anticipation of perceptual changes depending on motor signals. Mechanisms adopted from Dynamic Programming allow for behavior planning. Secondly, I will reinvestigate the growth and self-organization of a state space in the more rigorous framework of Hidden Markov Models. The goal are mechanisms that allow to organize a suitable latent state space also in the case of partial observability.

11/10/03 Binary Synapses and Fast Learning

David Barber

Binary synapses have been experimentally observed to exist in some neurons. Theoretically, they have been suggested as a possible solution to the problem of stablising memories (Stefano Fusi). An interesting issue is how learning may occur in these systems, since gradient based approaches would appear to be non-applicable. Recently Walter Senn proposed a `slow' stochastic learning mechanism. I'll present an alternative `fast' learning algorithm which, certainly for sparse memories, converges rapidly. A somewhat tongue in cheek suggestion is that this may provide a motivation for binary synpases in facilitating short-term memory.

04/10/03 Clever jelly: tricky problems in intracellular modeling

Fred Howell

There is increasing enthusiasm for looking beyond the genome > to systems issues (what does all this sequence, gene expression > and proteomic data mean? how do we make working models > of interacting networks inside a cell?). But there are a number > of characteristics of the activity inside a cell which make modeling > tricky, and new techniques need to be developed to address them. > In this talk I'll outline where existing techniques fall down, and > what characteristics are needed in future systems models.


Mark van Rossum

1) The breakdown of information theory for sparse signals. Evidence from the retina.

2) Transmission of population codes with spiking neurons; another reason for centre-surround circuits.

21/10/03 Clusters, Computation and Complexity.

David Willshaw

ESPRC have a new initiative that is funding clusters of like-minded scientists to get together to prepare grant applications under the theme of Novel Computation.

In this talk, I will review

(1) why this is happening.

(2) the scientific scope of the clusters.

(3) what being in a cluster is all about.

(4) what is in it for us/you.

(5) any deep truths that I may have learnt through deconstructing the background material to this initiative.

14/10/03 Eye fixation behaviour and the hemispheres

Richard Shillcock

When skilled readers read text they produce characteristic patterns of fixations locations and durations. We can model these behaviours by quantifying the uncertainty in each hemifield, as suggested by an anatomically based theory of visual word recognition based on the fundamental observation that the human fovea is precisely vertically split and projects contralaterally to the visual cortices.

07/10/03 Designing Tomorrow's Category-Level 3D Object Recognition System

Chris Williams

I'll give an overview of the workshop "Designing Tomorrow's Category-Level 3D Object Recognition Systems" held in Taormina, Sicily 8-10 September. In particuar I'll mention work by Yann LeCun (Invariant recognition of generic objects from shape), Kevin Murphy (Using the forest to see the trees: a graphical model relating features, objects and scenes), Rob Fergus (Object class recognition by unsupervised scale-invariant learning), Shimon Ullman (Fragment-based recognition and segmentation), David Lowe and Cordelia Schmid (Invariant local features for recognition) and Song-Chun Zhu (Markov chain methods in visual computing).

23/09/03 Testing for spatial dependencies in the positioning of two types of neuron

Stephen Eglen

Even in a small patch of neural tissue, many different cell types can be found. Are there any spatial dependencies in the positioning of a cell with respect to cells of a different type? Such dependencies may give us clues about how the different cell types develop. In this talk I will present results looking for spatial dependency in the mammalian retina, using a cross-correlation technique.

After the coffee break, for anyone interested, I will give a brief overview (or maybe Question and Answer session) of the R computing environment (

16/09/03 Behavioural specificity: Population level effects in a generalist biocontrol agent

Dean Baker

Insect foraging behaviour is a complex process in which a range of steps are recognized, and any change in the host specialization of an insect will depend on changes in one or more of these components. The parasitoid insect, Diaeretiella rapae, is a species that has evolved to attack aphid hosts exclusively, however evidence suggests that in some parts of the world, its host range has become restricted so that only a subset of available species are attacked. This talk concerns the identification of divergent foraging behaviour in an Australian population of D. rapae using a range of quantitative and molecular genetic techniques. Low genetic variation in this population could limit its evolutionary potential, including the ability to mount a response to newly introduced hosts, like the Russian Wheat Aphid, a species of considerable threat to Australia's grain industry.

09/09/03 Uncertainly in Artificial Intelligence 2003.

Amos Storkey

A report-back about some of the developments presented in the UAI2003 Conference.

26/08/03 On Likelihood-Based Learning

Michael Schouten

Likelihood-Based parameter estimation is one of the most widely used (and abused) learning paradigms for iid data. Applying these methods to data with more complicated dependencies can be facilitated by using graphical models. In this workshop, I first discuss several simple examples from Economics that illustrate how likelihod-based approaches are commonly employed in practice. I then present some results from my current research, which uses Bayesian Networks to estimate haplotype frequencies from samples that exhibit a variety of phylogenic structures.

12/08/03 Asymmetric Neighbourhood Effects and the Split-Fovea Model: left-handed English Readers

Alex McCauley

Orthographic neighbours (N) are defined as the number of same length words differing from a target word by one letter. For example, 'marsh' has two neighbours, 'harsh' and 'march'. N effects are robust and generally show facilitation effects for words with many N. A possible explanation is that a written word can activate lexical entries for 'similar' words, influencing recognition speed

An aysymmetric N effect (aNE) has recently been established using a lateralized lexical decision task. The standard N effect is seen in the right hemisphere (RH) but not in the left hemisphere (LH) of right-handed native English speakers.

In the split-fovea model, visual word recognition is conditioned by the precise splitting of the foveal image about the fixation point. Letters to the left and right of fixation are projected contralaterally. The RH reflects processing of initial, more informative, letters. It may be adaptive for the RH to activate lead-N, while the LH does not activate end-N. If the RH applies this strategy when performing lexical decision on the whole word, we would find the aNE.

Left and right-handers can differ in language lateralization and in hemispheric asymmetries for lexical processing. However, with the same reading exposure as right-handers, left handers should not differ in their expression of the aNE. Here I report a study which addresses this hypothesis.

05/08/03 Template matching, matching primitives and learning object parts

Michalis Titsias

The talk is split into two parts. In the first part I will review some well-known approaches for object recognition giving emphasis on model-based vision, where the recognition process is based on matching an appearance model of the object class (e.g a simple template) into a test image. At the second part of the talk I will describe a probabilistic model for learning panoramas in video sequences with a moving camera.

29/07/03 July 29th, 2003

Rebecca Smith

Mitral cells are the main projection neurons of the olfactory bulb where it is thought a large part of olfactory processing takes place. Thus it is likely that they play an important role in the detection and recognition of odours. The mechanisms involved in the regulation of the firing patterns of mitral cells are poorly understood therefore in vivo recording experiments and computational modeling techniques have been employed to provide a clear answer to this question. I will discuss all the cells found in the olfactory bulb and the role they could play in the regulation of mitral cell firing. I will cover the mitral cells in more detail to demostrate how mechanisms internal to the cells are also involved.

22/07/03 ANC Workshop - July 22nd, 2003

Paulo de Castro Aguiar

My talk will be divided in three parts:

1)First I will show the final results and conclusions of the work that I have been doing on hippocampal mossy fibre boutons as dynamical synapses.

2)Secondly, and trying to follow the ideas proposed in the "ANC Day" for our workshops, I will discuss a paper on spatial learning, a interesting subject that is not very usual to ear about in our workshops. The paper is "The involvement of recurrent connections in area CA3 in establishing the properties of place fields: a model", by Ka'li and Dayan. The motivation of this paper is to create an attractor network model of area CA3 in which local, recurrent, excitatory and inhibitory interactions generate appropriate place cell representations from location and direction specific activity in the entorhinal cortex. I will give the necessary basic background to understand this work.

3)Finally, I will show how, using a webcam and basic techniques of machine learning, you can use basic gestures to send commands to your computer.

15/07/03 Aspects of Mutual Information Processing

David Barber

The reliable transmission of Information has often been proposed as a central goal of information processing in both natural and artificial systems. I'll cover some basic aspects of Mutual Information and discuss applications in a variety of systems, from data transmission to population coding.

08/07/03 A Grand Theory of Dopamine

Andrew Gillies

The neurotransmitter dopamine is strongly implicated in many dysfunctions of the nervous system. From Parkinson's disease to schizophrenia to obessive compulsive disorder to drug addiction. What is it doing in the normal brain?

I will present a recent review of dopamine in the normal brain, and introduce an initial stab at a grand theory of dopamine; i.e. what is the role of dopamine in normal function, and how can perturbed dopamine levels be involved in so many disorders?

01/07/03 Converging fMRI, ERP and ERF reveal the role of the left temporal lobe in synatactic processing

Martin Meyer

Traditionally, based on neuropsychological obsevations Broca's area in the left frontal gyrus has been described as housing synatatic functions (Grodzinsky 2000. Recent auditory event-related brain potentials (ERP) studies examining the processing of syntactic violations provide evidence that indirectly support this view (Hahne & Friederici 1999) as they report a larger negative going waveform (ELAN) recorded from electrodes over left inferior frontal sites following syntatic violations. However, due to the 'inverse problem' one cannot infer from the topographical distribution of ERPs the corresponding neural source. In addition, an fMRI-study in which particpants were presented with the identical set of sequences identified a region in the left anterior superior temporal region (planum polare) as being particularly sensitive to syntactic anomalies (Meyer et al. 2000). An event related magnetic field (ERF) analysis based on MEG recordings taken during a separate session of this experiment was performed. This analysis aimed to localizing the neural source of the ELAN more precisely as this technique provides a better spatial resolution than EEG. Therefore a dipole source localization using a realistically shaped standard volume conductor model and fMRI coordinates on inferior frontal and anterior temporal sites provided by the previous fMRI-study, were selected as the seeed points for constrained dipole fitting (Friederici et al. 2000). The results clearly demonstrate that left temporal planum polare rather than Broca's area supports syntactic processes. This finding indicates that the traditional view on the function of Broca's area in speech processing may need to be revised.


Y. Grodzinsky, Behav. Brain Sci.23, 2000.

A.D. Friederici, Hum. Brain Mapp. 11, 2000.,

A. Hahne, J. Cog. Neurosci. 11, 1999.

M. Meyer, Cog. Brain. Research 9, 2000.

24/06/03 Reading words aloud: not quite the same with one- and two-syllable words

Marielle Lange

Theories of print-to-sound conversion developed around computational models have generally left polysyllabic words out of their explanation because of the difficulties to represent them in the system. As a result, even though they have been specified in full detail and shown to produce patterns of performance similar to those of skilled readers with monosyllabic words, none of them can actually be considered as complete. Here, I introduce a linguistic description of the grapheme-phoneme relations (i.e., othographic units mapped onto a single speech sound) in disyllabic English words and discuss the way this description challenges some of the hypotheses derived from the properties of monosyllabic words. In particular, I will comment upon the preponderance of the information about stress placement for correct pronunciation assignment, the difficulty to appropriately segment the string into graphemes when disyllabic words are introduced (e.g.. AI in WAIVE vs NAIVE), the contribution of a knowledge of the neighbouring letters for correct translation (e.g., C in CELL vs CALL),and the importance of the consideration of the phonetic properties of the adjacent letters for appropriate generalisation of the regularities (e.g., SS is pronounced as SH before a yod, in MISSION, PRESSURE, OR MISS YOU pronounced in rapid order).

17/06/03 Adaptive Cognitive Modelling: Beyond PDP

Chris Williams

In this talk I will present the paper "Mechanisms of long-term repetition priming and skill refinement: A probabilistic pathway model" by Mozer, Calagrosso and Huber, which uses a Hidden Markov Model (HMM) to model long-term repetition priming. The paper is to be presented at the 25th Annual Conference of the Cognitive Science Society and is available from

If there is time, I will also mention recent work on Extreme Components Analysis (XCA; joint work with Max Welling and Felix Agakov), and on a conjecture on the number of modes of a Gaussian mixture (joint work with Miguel Carreira-Perpignan).

27/05/03 The tortoise and the hare: two approaches to working out the function of cortical microcircuits

Fred Howell

One approach to brain modeling is to wait until we have sufficent experimental data so that we can build data driven models which can be proven to represent reality. Another is to suggest that we already know enough about the dynamics of neurons to be able to start to build speculative models of how we think it ought to work. In this talk I'll present how one might gather the disparate experimental data to help with the first approach, and give some speculations on the second.

20/05/03 The Corpus Callosum and word recognition

Richard Shillcock

I will review some of the literature on the development, normal functioning,and impairment of the corpus callosum and discuss some of the implications for the modelling of visual word recognition.

13/05/03 Light and informal presentation of Catacomb 2.0

Paul Rogister

A software for biological modelling developed by Robert Cannon, and of my work on an XML-based external component system for Catacomb.

06/05/03 Beyond Excel: Axiope Software for Data Management and Data Sharing

Nigel Goddard

The growing amount and diversity of biological data needs new software tools to help scientists organise, catalog and share it. There are limitations to how far spreadsheets like Excel can be pushed in this direction; and there is a limit to how much information can be encoded in a file name. Database programs are useful in many contaxts, but are typically too complicated to set up and maintain for many lab applications.

We are developing a suite of tools for dealing with "metadata" about experimental data, including the possibility of sharing it within a lab and on the web, and making such data available for furture "data mining" experiments. The starting point for the Axiope approach is the ''Catalyzer'' program which lets you create catalogs of data. It is more powerful than spreadsheets like Excel for data management as it allows tree-structured data as well as tables. It is as flexible as the file system in coping with any type of data, but allows arbitary amounts of structured descriptions of the files: you don't need to pack everything into a filename. Axiope catalogs have a key advantage over other forms of data - unlike spreadsheets or databases it is simple to combine any number of catalogs to create a merged view of all the data. This makes it possible to present a browseable index of all the experimental data produced by a single researcher, lab, consortium, or even scientific discipline.

29/04/03 Biological Networks in the Mammalian Synapse

Douglas Armstrong

What is this systems biology thing all about? I will attempt to introduce part of the field and briefly review some recent research into biological networks and their properties

Over the past 12 months or so we have been working with Seth Grant's group in Neuroscience on a protein interaction database system. The core data revolves around proteins in the post synaptic density in the mouse brain. The development of the database will be described briefly. The rest of the talk will focus on the analysis of the protein networks that have emerged from the dataset.

15/04/03 Foresight: Cognitive Systems

David Willshaw

am involved in the 'Foresight: Cognitive Systems' programme which is run by the Department of Trade and Industry for John Taylor, Chairman of the UK Research Councils. The programme is concerned with trying to identify profitable areas of research by bringing together scientists to discuss possible new areas of research, in this case at the intersection of the life sciences and the physical sciences.

My talk will have three parts:

1. An introduction to the Foresight: Cognitive Systems programme

2. A review of the briefing paper entitled 'Self-organisation in the nervous system' which I prepared for this programme.

3. Information about how people can get involved in this programme.

08/04/03 FFT method for missing data

Amos Storkey

Discrete Fourier transforms and other related Fourier methods have been practically implementable due to the fast Fourier transform (FFT). However there are many situations where doing fast Fourier transforms without complete data would be desirable. For example someone might wish for fast methods of Fourier analysis of textures of non rectangular image patches or for dealing with missing data, such as music data from scratched compact discs or data with regular interruptions. The FFT is undetermined in these situations. In this paper it is shown that the FFT algorithm can be formulated as a belief network, thereby allowing suitable priors to be set for the Fourier coefficients. Furthermore efficient loopy propagation methods between cliques of four nodes allow the Fourier coefficients to be inferred and the missing data to be estimated. This method is compared with the common approach of setting missing data to zero or to interpolation. It is tested on generated data and for a Fourier analysis of a damaged audio signal. The approach is also suitable for use in situations where data is not exactly some power of two in length.

01/04/03 Place Field Formation in CA1

Kit Logden

Models of place field formation in the hippocampus use the recurrent connections in area CA3 to create attractors of spatial location. Area CA1 recieves the majority of CA3 output and has the sharpest place fields. However, it was shown last year that the direct cortical input to CA1 maintains the place fields when CA3 is lesioned. This is exciting because it indicates that the cortical input can bypass the hippocampal loop to recall hippocampal representations, or that it controls the coding in CA1.

My model of place field formation in CA1 uses balanced inhibition to ensure that each cell responds to correlations in its cortical input. The effect of plasticity in the model is to broaden and move the place fields. This is consistent with studies of the effect of experience on place fields. However, studies of plasticity in CA1 paint a complex picture that I will argue supports the control of CA1 coding by direct cortical input.

25/03/03 Quantifying nature and nurture

Michael Schouten

The objective of this talk is to provide an introduction to my research area for those without a background in quantitative genetics. I will begin by discussing Mendelian inheritance and will illustrate how graphical models are used in Genetic Counseling and Risk Assessment for human disease. I will then outline more complex methods of inheritance and the role of molecular markers in population genetics. I conclude with a summary of my research topic, which entails using MCMC sampling on Bayesian Networks to establish an association between genetic markers and quantitative traits.

11/03/03 The distribution of synaptic weights

Mark Van Rossum

The response of a neuron is shaped by the strength of its synaptic connections. Using spike timing dependent plasticity (Markram et al. 1997, Bi and Poo 1998) we calculate the evolution of the synaptic weight distribution. The relative amount of depression is independent of the initial synaptic size, in contrast, potentiation gets smaller for larger synapses. This leads to a stable weight distribution of synaptic weights, which is calculated exactly and is similar to experimentally observed distributions of quantal amplitudes. Competition, observed earlier in similar models, is almost absent, but is restored by including activity dependent scaling (Turrigiano et al. 1998).

Some people might note some overlap with older work of me. This is not incidental! Hopefully, there will be some new results as well.

25/02/03 Sequential Learning of Mixture Models

Michalis Titsias

Consider the unsupervised learning scenario: Given some data find a summary or description of that data. The sequential learning can be thought as trying to build that description sequentially focusing on different aspects of your data at a time. Alternatively we can think of this as trying to fit the unsupervised learning model sequentially. For example, it is known that PCA can be carried out in this fashion by learning the principal components one after the other.

We present an algorithm that trains a Gaussian mixture model sequentially; learning one Gaussian at a time. In addition, an algorithm that we have developed and learns multiple objects in images (see Chris Williams previous talks) is also a sequential algorithm of this kind. We show also some demos of experimental work.

18/02/03 Outside and inside schizophrenia:
computational models and subjective experiences

Nicola De Pisapia

Schizophrenia is one of the most challenging and mysterious brain disorder. I review some cognitive and computational studies, and then contrast them with the issue of what is it like to be schizophrenic. While modelling can improve the understanding and medical treatment of this disorder, the sympathetic approach can help in understanding the persons who are affected, in reducing social discrimination, and even in giving important insights into the human culture in general.

11/02/03 Dynamic Synapses and Tuning Neurons

Paulo de Castro Aguiar

It's undeniable the ability of the nervous system to perform temporal information processing: tasks like distinguishing two brief tone pulses separated by either 50 or 100 ms, coordinating arm movements or playing music are examples of temporal pattern interpretation and control of temporal pattern production that remain poorly understood in terms of their underlying mechanisms.

In this talk I will speak about dynamic synapses and how this (temporal processing) mechanism can be used to perform temporal discriminations. Recent work by Buonomano, Markram and Tsodyks will be presented as well as some results that I've obtained using realistic cell morphologies and synaptic parameters in NEURON simulations. At the end I plan to give a simple view of my PhD project and explain why dynamic synapses play an important role on it.

04/02/03 Optimal Hebbian Learning : a Probabilistic Point of View

David Barber

I'll describe some recent work with Jean-Pascal Pfister and Wulfram Gerstner that aims to derive a spike time dependent learning rule from a probabilistic optimality criterion. The method is an extension to continous time of the discrete Markov framework I've previously presented inwhich the idea is to create a model in order to obtain quantitative results in terms of a learning window. This is achieved by maximising a given likelihood function with respect to the synpaptic weights. The results are consistent with the actual knowledge of back-propagating action potentials.

21/01/03 The Mysteries of Dopamine

Andrew Gillies

Until the 1950's dopamine was not considered to have a functional role in the central nervous system, rather it was thought to be simply involved in the biosynthesis of noradrenaline and adrenaline. Now, dopamine is thought to have roles in Parkinson's disease, reinforcement learning, drug addiction, Attention Deficit Hyperactivity Disorder, affective disorders, schizophrenia and much more. It is now one of the most studied neurotransmitters.

I will present a overview of material given at the recent Society for Neuroscience (Nov, 2002) conference highlighting a new view in the role of dopamine in drug addiction. I will also touch on some of the other areas dopamine is involved creating links between proposed dopamine functions.

14/01/03 Two quirks of evolution

Marcus Frean

It is natural to think of competition among species as giving rise to hierarchies, yet it turns out that evolution and spatial structure can combine to make cyclic competitions (like the game of rock-scissors-paper) surprisingly stable. We predict that, paradoxically, in such scenarios the species which is the worst at competing is least likely to go extinct - in effect, 'survival of the weakest'.

It's also natural to think of learning as being evolutionarily advantageous, but is the ability to change behaviour quickly necessarily a good thing? I'll describe a Prisoner's Dilemma scenario in which adaptability is disastrous, namely two individuals who interact and learn about each other, but at different rates - in this case the slower learner comes out substantially on top. As an application of this idea, two symbiotic species can 'learn' about each other over evolutionary time, which may explain why some endosymbionts seem to get such a raw deal while their hosts reap nearly all the benefits.

17/12/02 The Neural Correlates of Facial Attractiveness

Adam McNamara

Faces are amongst the most important stimuli for human social functioning, and have been described as the 'single most important pattern in our environment' (Ellis and Young, 1989). As such, they have been the focus of much research from cognitive, ethological and physiological perspectives. Integration between such approaches, however, has often been lacking. Although brain areas involved in face processing have been identified, little work has considered variability within these regions in response to variability in face stimuli (with the notable exception of perception of emotion, see Young, 1998, for review). Theories from evolutionary biology have stimulated behavioural research investigating the characteristics of facial attractiveness, employing experiments that manipulate face stimuli along dimensions that may have biological relevance (such as symmetry or averageness, see Thornhill and Gangestad, 1999, for review). However, the neural correlates of attractiveness judgements in such experiments remain unexplored. Facial characteristics that are known to influence behavioural responses and that are implicated in biological and cognitive accounts of face processing seem a practical starting point to investigate the neural correlates of attractiveness judgements. One such characteristic is the 'prototypicality' or 'averageness' of a face - a concept central to many accounts of facial perception. Faces closer to average facial proportions are harder to recognize, yet paradoxically more attractive (Young, 1994; Langlois & Roggman, 1990; Penton-Voak & Perrett, 2001). The aim of this experiment is to investigate the neural correlates that support the perception of faces that vary in 'averageness'. Mixed Block/Event design fMRI procedures were employed and preliminary analysis shows differential activation between attractive and unattractive stimuli as judged subjectively by the subjects in the scanner in areas such as the fusiform gyrus, hippocampus and reward regions of the midbrain. This study is in collaboration with Dr. David Donaldson and Dr. Ian Penton-Voak of Stirling University.

03/12/02 "What makes a good kernel?"
"Harmonising Chorales in the Style of J S Bach"

Chris Williams

The first is a practice talk for the NIPS kernel machines workshop. The second is a brief description of some work done by Moray Allan as his MSc project (2002) on using HMMs for harmonising chorales in the Style of J S Bach.

10/12/02 Language differences in the two hemispheresof the brain

Richard shillcock

There are written language processing differences between the two hemispheres of the brain. To what extent are they the result of innate predispositions, as opposed to being learned? I will present some joint work in which connectionist modelling and human experimentation suggest that the written language to which the brain is exposed may condition the two hemispheres to behave differently in lexical processing.

26/11/02 Cartoon Modeling of proteins

Fred Howell

In this talk I'll present some work Dan Mossop has been doing building a "cartoon level" simulator of protein interactions. Molecular biology books typically use static cartoon descriptions to hypothesise underlying mechanisms for pathways and interactions of proteins.

This simulator represents proteins as simplified 3D cartoons with structures and bonding sites derived from a mixture of protein structure data (from PDB)and models derived from cartoon descriptions.

The idea is to fill a model cytoplasm with a menagerie of different protein building blocks (with specific bonding sites, states and probabilities and then see what happens over time.

The examples to date include: - formation of part of the post synaptic density from constituent proteins - formation of spherical vesicles from clathrin - vesicle transport using a motor protein - formation of microtubules

I think there are many other potential applications of this level of modeling, such as: - spontaneous assembly of viruses - modelling the flagellar rotary motor complex of a bacterium - a detailed model of ribosome formation and operation

And getting more ambitious: - a molecular model of cell division

12/11/02 Odour Coding and Jacobsons Organ

Rebecca Smith

Basically the talk will be split into two halves. In the first I'll talk about the main theories of odour coding and how biological simulations can be used to support these, and in the second I'll talk about Jacobsons organ and the accessory olfactory system, just because i find it interesting.

05/11/02 Building efficient interoperability for
neuroscience simulations components.

Paul Rogister

In this talk i will review our efforts, as part of the NeuroML project, to solve the persistent problems users face when trying to reuse other people's models or simulation's packages within their own. Our approach, based on XML, is targeted at two levels : the model's description and the simulation software. Part of it has already been discussed in precedent talks, exept the most recent development : an XML based system designed for module interoperability.

29/10/02 Recent developments in the study of functional connectivity of the Hayling task in the High Risk project

Enrico Simonotto

Today I will discuss some recent developments in the study of functional connectivity of the Hayling task in the High Risk project

22/09/02 Flies in Space

Douglas Armstrong

In preparation for two invited seminars in Novemeber I would like to subject you all to a trial run of my latest edition of 'Flies in Space'. Previous and on-going work into dissecting the neural pathways that underly gravitational behaviour in insects will be described.

If time permits I will also review this summer's workshop on Neuroinformatics in Model Organisms.

Since ANC has changed quite a bit, I will also briefly mention some other on-going work.

15/09/02 Associative Networks for Recognition of Familiarity

David Willshaw

Neurons in the perirhinal cortex are known to be able to discriminate between familiar and novel visual stimuli. How the familiarity/novelty distinction is computed is not known but recent work has suggested that this distinction is mediated through the operation of an associative network.

There has been extensive work on the performance of associative networks in retrieval tasks; how they could be used in a recognition task has not been examined in any detail.

In this talk, I shall review some work carried out by Bogacz & Brown (NETWORK, 13, 457-485) on their theory of how linear associative networks can be in this recognition task. They review some fairly complicated mathematics, showing that (1) the performance of their model networks for a recognition task is much greater than for a retrieval task; (2) some learning rules are better than others, a covariance-type rule seeming to be the best; (3) for the learning rules that they investigated, performance in the recognition task does not depend significantly on the so-called sparseness of encoding of information in the patterns being stored.

I shall examine each of these claims and shall then review some recent mathematics (in progress), towards understanding whether for this task there is indeed an optimal learning rule and how it relates to the optimal rule derived for the retrieval task (Dayan & Willshaw, 1991. Bio Cyb, 65, 253-265).

08/10/02 A scattering of problems and solutions
.And hopefully some of them match

Amos Storkey

As it is the beginning of a new year, with many new faces, after a quick bit on where I am coming from (translation: sermon on Bayes), I will give a quick overview of some of the problems I am involved with, giving a little more detail about the problem of record linkage. I also hope to get some sort of discussion going about where probabilistic models might help in the type of work going on in anc.

27/08/02 Fast propagation and computation in layered
feed-forward networks

Mark van Rossum

One of the classic neural coding schemes is rate coding. In rate coding information is coded in the firing rates of the neurons. This idea comes from the experimental observation that presenting the preferred stimulus often leads to an increase in the firing rate of a neuron. However, rate coding has been criticized because given the typical firing rate (up to 100 Hz) and given the Poisson-like variability in typical spike trains, substantial averaging time is needed to estimate the signal. Experimental data suggests there is only very little time for such averaging, as humans can categorize complex visual scenes within 150 ms.

We argue that under the right conditions, computation with rate coding is fast. First, we study how firing rates propagate through a layered feed-forward network with 20 integrate-and-fire neurons per layer. All neurons are primed with an independent noise current and a net excitatory current. Time varying stimuli can propagate rapidly and accurately through many layers. Next, we demonstrate how elementary computations are implemented, and show what the fundamental limits to the computational speed are. Finally, we build a motion detecting network with these networks.

13/08/02 The Mathematical Brain

Tim Hely

In a couple of weeks time I will be leaving ANC to pursue pastures new. I will be starting a 1 year PGCE at Moray House to qualify to teach Physics and Maths to secondary school children (sounds like fun!) To link two of my interests, I will be summarizing Brian Butterworth's impressive book "The Mathematical Brain" along with some other mathematical bits and bobs thrown in. I will discuss the evolution of counting and the many different ways in which this is done around the world. A brief history of numbers from zero to infinity. And answer questions such as "Why is D the Roman numeral for 500? Do babies and infants have innate numerical abilities? How to be a mathematical prodigy (it may not be too late!) What parts of the brain are wired for numbers? What *does* subitizing mean?" I will also discuss a number of clinical case studies of people who have either no real concept of number or who have lost their numerical ability as a result of developmental dyscalculia or brain damage.

Almost as much fun as you can have in a room with other people.

06/08/02 Synaptic plasticity and how to model it

David Sterratt

This talk is intended as an introduction to synaptic plasticity at functional and molecular levels. I will focus on long-lasting synaptic changes and ways of modelling them.

It's a draft version of a talk I'll be giving to the summer school on Neuroinformatics simulation tools in September, so comments would be most welcome.

30/07/02 Modelling Neurons. The Black Arts

Andrew Gillies


23/07/02 Elements of Commercialisation

Nigel Goddard

For the last few months I have been working with Fred Howell and Robert Cannon to put together a spinout company to commercialise some of the scientific software we have developed. In this short talk I will review what we have learned about the requirements to be successful, the process of commercialisation, and the direction of Government policy.

16/07/02 The Dorsolateral Prefrontal Cortex:
it's role in behaviour and a computational model

Nicola De Pisapia

I'll review experiments on three basic processes: active maintenance in Dorsolateral Prefrontal Cortex (Fuster, Goldman-Rakic, Watanabe), chunking in the Basal Ganglia (Graybiel), and the shift of neural activity from the front to the back of the cortex during motor skill acquisition (Shadmher). I'll then use the interpretations of these experiments to justify my connectionist model on how the primate brain can achieve and improve its Planning capacity. I tested my views within a Reinforcement Learning framework, and I'll show some results from several simulations.

09/07/02 Learning in Spiking Neural Assemblies

David Barber

We introduce a general statistical framework for learning in a large class of biologically plausible models of a network of spiking neurons. We apply our framework to the derivation of local learning rules for learning temporal sequences for a phenomenological model of spiking neurons and show that the rule qualitatively matches empirically observed temporal asymmetries. We further show how to include mechanisms such as synaptic depression, and outline how the framework is readily extensible to learning in networks of Hodgkin-Huxley spiking neurons. We discuss models of quantal release mechanisms, and show that in such cases, a framework for learning is more complex.


02/07/02 Some highlights from ECCV 2002

Chris Williams

I present some highlights of statistical vision work from the European Conference on Computer Vision 2002, notably Object Recognition as Machine Translation (Duygulu et al), Learning Montages of Transformed Latent Images (Pal et al), What are Textons? (Song-Chun Zhu et al). I will also describe a NIPS submission (joint work with Michalis Titsias) about Learning About Multiple Objects In Images: Factorial Learning without Factorial Search.

25/06/02 Catting Art Galleries
- an approach to scientific data integration?

Fred Howell

The web in 2002 looks like a huge document, parts of which are generated on the fly from databases, and much of which is indexable by free text search on Google. Will the web in 2003 look like a huge database? In this talk I look at an approach to building a federated web database of all online collections in all art galleries, which would make browsings like "show me thumbnails of all works by picasso and matisse sorted by date" a few clicks of a web browser rather than a database programming effort. The approach also has applications to scientific data.

18/06/02 fMRI responses to non-linguistic prosody

Martin Meyer

Brain imaging techniques like fMRI provided evidence that the brain's right hemisphere is particularly processes linguistically relevant prosodic information during speech perception and production. This might raise the question of whether the RH is also adept at subserving non-linguistic prosody. Data obtained from three fMRI studies will be presented which point to a prevalent role of the RH in processing non-linguistic prosody. The the first experiment is aimed at identifying brain regions which subserve selectively in affective rather than linguistic prosody. The second experiment used human laughter and isochronic environmental sounds such as non-linguistic stimulus to examine the cerebral substrates of rhythm. The results obtained from the third experiment suggest a greater sensitivity of the RH in recognizing artifically manipulated relative to natural voices. In summary the three experiments point to a particular involvement of right supratemporal areas in processing non-linguistic prosody.

28/05/02 Talk today

Rebecca Smith

I'm talking about mitral cells again but I'm going to be concentrating almost entirely on the progress I've made with modelling them so hopefully you'll find that interesting. I've just rencently come up against a problem which I'm having difficulty seeing my way around so I would be extremely grateful for as much feedback about that if possible.

14/05/02 All you ever wanted to know about Molecular Genetics
but were afraid to ask...

Douglas Armstrong

In the light of the two current reading groups I thought it might be a good time to go over some basic molecular biology for those that are interested. As its 'in-house' I can go over the basics, starting with genes and proteins and working up to microarrays and proteomics.

07/05/02 The roles of Chemoaffinity,
induction and ephrins in the formation of nerve connections

David Willshaw

The problem that I address concerns how ordered patterns of nerve connections are developed in the nervous system. The talk has three parts.

1. I will review some of the experimental evidence, the arguments for and against and some dilemmas concerning the idea of chemospecificity put forward originally by Roger Sperry in the 1940's. I will focus on the developmment of connections in the retinotectal system in lower vertebrates.

2. I will then summarise the Tea Trade Model of induction of chemical markers proposed by myself and Christoph von der Malsburg in the late 1970's.

3. In the final part of my talk I will discuss the new evidence concerning how the interpretation of the Eph receptors (in the retina) in combination with the associated ephrin molecules (in the tectum) as the basis for map-making according to chemoaffinity has spawned a new set of models and I will suggest tentatively how the induction model can be applied to new experimental findings.

This talk is a substantially expanded version of my presentation at the ANC getting-to-know-you day.

30/04/02 Menu for today's talk

Amos Storkey

Starter: Goldbach conjecture soup For over 150 years Goldbach conjecture has remained unproved, and yet number theorists are pretty sure it is true. This will be a quick hand-wavy ditty on why they might think this as an illustration of how probabilistic arguments can aid intuition.

Main Course: Satellite track steak with renewal string vegetables A fifteen minute review of work on detection of spurious objects in astronomical databases, with an introduction of the renewal string method.

Dess1ert: Super-resolution and cream A simple approach to enhancing image resolution using dynamic tree methods which is (perhaps, maybe) moderately okay :-), and produces centre-surround effects in the latent layer.

Coffee: Downstairs in the common room

23/04/02 Simple Sequence Learning in Noisy Recurrent Neuronal Networks

Tim Hely

I will discuss work in progress on carrying out a short term memory task using a biologically plausible recurrent neuronal network architecture. In contrast to most traditional neural networks, the network I am using has sparse, asymmetric connectivity patterns and a low background rate of firing activity. The task involves storing a short sequence of numbers and is similar to remembering a telephone number. In general the network copes well with "easy", random sequences (e.g. 1,4,2,7 ...) but can also learn "hard", non-random sequences (e.g. 2,2,2,4,2,2,6) with a memory window of only one timestep. Key aspects of the problem involve how the input/output patterns should be encoded/decoded and what learning (and forgetting) rules seem to give best results. I will also discuss the problems of recall without a prompt, category saturation and continuous local, online learning.

16/04/02 Converting Runit to Neosim

Paul Rogister

A short overview of the first results in converting the Runit simulator, a temporal sequence learner, to the Neosim module-based framework and the NeuroML interfaces. Nothing extraordinary yet, but some nice picture of the easy-to-use NeuroML graphical component that any NeuroML compliant modules can include. A real boost to simulation coding?

09/04/02 Hopfield Nets, Temporal Sequences and Markov Chains

David Barber

Hopfield networks are models of distributed computation, of interest to both theoretical neurobiologists and physicists. The special case of Synchronous Hopfield networks can be viewed as Markov Chains. I'll discuss a simple learning algorithm for this case that outperforms Hebbian learning for storage/reconstruction of static and temporal sequences.

02/04/02 (Extra-Hebbian) neuronal plasticity and how we should model it

David Sterratt

Hebb (1949) suggested that the synapse between a pre- and postsynatpic neuron increases in strength due to the presynaptic neueron taking part in causing the postsynaptic cell to fire. The term `Hebbian' now broadly referrs to plasticity that is local to the synapse connecting two cells considered to be active.

Evidence from the hippocampus dating back to the early 1970s supports this hypothesis. However, more recent evidence suggests that synaptic plasticity in the hippocampus also has an extra-Hebbian component, i.e. that synapses other than those connecting active pre- and postsynaptic neurons are potentiated or depressed. In the first part of the talk I will present some of this evidence.

One question these results lead to is "what is the biophysical basis of synaptic plasticity?". In the second part of the talk, I will summarise the (or rather "my") current understanding of synaptic plasticity in the hioppocampus and approaches to modelling it.

26/03/02 Functional MRI and functional connectivity of
the Hayling sentence completion test in subjects at high risk
of schizophrenia: preliminary results

Enrico Simonotto

The Edinburgh study of subjects at high risk of schizophrenia is one of the first large longitudinal studies of the development of the disease which makes use of fMRI with the aim of examining the relationship between disordered neuronal activity and the development of the clinical features of schizophrenia

We have nearly completed the first phase of the study. I will discuss the preliminary results of functional localization and functional connectivity on a rather large (n=79) group of participants in the study. The participants in the study are split in three groups: controls (n=21), high risk subjects without symptoms (n=37) and high risk subjects with symptoms. While behavioral and functional localization results show little differences between groups, functional connectivity results point to large differences in the BOLD signal in the high risk with symptoms respect to the controls.

19/03/02 Top down theories of the Basal Ganglia

Andrew Gillies

My own work focuses on bottom up/data driven approachers to understanding the nervous system (in particular nuclei of the basal ganglia). It is good from time to time to step back and review the top down views which address the more global questions, such as "what is it exactly that the basal ganglia do?". There are currently two main top-down theories of the basal ganglia, that although in competition with each other share much in common. I will review these top down theories, their origin's, and the evidence for and against them.

05/03/02 Modelling the bilateral distribution advantage effect

Padraic Monaghan

(Work with Stefan Pollmann, Day Clinic for Cognitive Neurology, University of Leipzig, Germany)

How do the two hemispheres of the brain communicate and cooperate in solving cognitive tasks? This question has recently benefiitted from implemented computational models of cognitive functions (e.g., Reggia, Goodall & Shkuro, 1998; Jacobs & Kosslyn, 1994).

We report here a neural network model of a task that has been widely used to investigate hemispheric processing. When subjects are required to match the shape of letters, then presenting both letters within one visual hemifield results in quicker and more accurate responses than when letters are presented one to each visual hemifield. However, when the task is to match the name of letters (so, "A" and "a" would be a matching trial), then responses are quicker and more accurate when letters are presented one to each visual hemifield (Belger & Banich, 1998).

This effect, known as the bilateral distribution advantage (BDA), has informed psychological research into the nature of hemispheric processing, distribution of resources between the hemispheres, and hemispheric transfer of information. We here report a simulation of the effect, and relate the performance of our model to theories of interhemispheric communication.

26/02/02 Introduction to Gaussian process models

Matthias Seeger

Due to some demand, I'll give another introduction to Gaussian process models, and how they can be used for regression estimation and binary classification, but also how they may be used in more complicated models. GP models can be understand from two different angles, both familiar yet a bit different from traditional statistical models. I'll introduce both:


- process (or random field) view: Everything's Gaussian!

- parametric (or weight space) view: Everything's linear!

If there's time, I'll talk about a PAC-Bayesian theorem for GP classification.

12/02/02 Human visual information processing

Richard Shillcock

I will talk about what happens between visual information impinging on the human retina and contact being made with stored information in the cortex. I will try not to repeat too much from previous presentations on our split-fovea connectionist modelling work. The goal will be to elicit some interdisciplinary speculation from others in ANC regarding more mathematical approaches to data fusion, etc. I will talk about combining information from the two hemifields, suppressing inconsistent information, vergence, contacting visual templates in longterm storage, and maintaining a representation of the scene, among other things.

05/02/02 Learning about Objects with Transformed Mixtures of Gaussians

Chris Williams

Frey and Jojic (2001) introduced an architecture for transformation invariant clustering, so that images of the same object can be clustered, ignoring transformations such as translation, rotation etc. Recently they have extended this work so that object models can be learned when multiple objects appear in an image at the same time. I will describe this work and also recent work with Michalis Titsias about an alternative approach to the problem.

If there is time left at the end I will tell you about the APRIORI algorithm for finding frequent itemsets and association rules.

29/01/02 Split Architectures and Symmetry

John Hicks

Previously I have talked about split neural networks and how such architectures can be effective for processing visual information, noting that there are certain emergent strategies of processing particular to these split architectures. Later, it was shown that variations in the class of stimuli had a secondary effect on such strategies. One such variation involved palindromic (i.e. symmetric) stimuli. The final model, presented here, is further evolved to look at stimuli taken from behavioural studies on the perception of symmetry. I'll cover some behavioural/theoretical work on the perception of symmetry, including that of Bayliss and Driver, and then talk about a related model. Employing the split architecture networks, the psychological account is brought partially into question by effects that suggest much of what has been termed ``psychological'' may be in fact due to anatomical factors in the bi-hemispheric visual processor. Thus a simpler treatment of symmetry perception might be available than that provided by Bayliss and Driver.

22/01/02 Implementing state machines with spiking neural networks

Fred Howell

The idea of the cortex being composed of an array of canonical microcircuits has been around for a while - but there have been fewer ideas of what this canonical microcircuit actually does.

In this talk I'll present a speculation that the important behaviour of this circuit is to learn (and usually rapidly forget) state-machine type behaviour using short term dynamics of synapses as a transient memory.

15/01/02 Functional insights into hippocampal CA1 interneurons

Kit Longden

Damage to specific neurochemical classes of hippocampal interneuron is associated with epilepsy and Alzheimer's disease, amongst other pathologies. Recent research has classified hippocampal interneurons by morphology, electrophysiology, and neurochemistry, revealing an incredible diversity, and these classifications rarely overlap. The mechanisms of their putative functional roles could explain which characteristics are relevant, but these are little understood.

After introducing some of the interneurons, I will describe some modelling work on oscillations by Traub and Co. to introduce a current accepted understanding of CA1 interneuron function. Finally, I will discuss how my recent work implementing a divisive inhibition mechanism for an associative memory network can be integrated with, and advance, this understanding.

08/01/02 Neurocognition of speech evidenced by 3T fMRI

Martin Meyer

The cerebral organization of speech is still a subject of debate. A vast body of evidence supports the view that left perisylvian areas subserve lexical and grammatical aspects of speech comprehension, while earlier lesion studies and recent data obtained from functional neuroimaging point to an essential role of right perisylvian areas during discrimination of prosodic information in tonal sequences.

Here, data obtained from two recent fMRI studies on speech processing at the sentence-level will be introduced. The results demonstrate a strong involvement of distinct areas in the left hemisphere whenever lexical and syntactic information is emphasized but challenge the widely held notion of Broca's area 'housing syntactic operations'. Further, the data provide strong evidence that right cortical areas particularly subserve pitch modulations (sentence melody). A bifrontal network including the fronto-opercular cortex might serve as an interface synchronizing prosodic and syntactic information during speech processing.

A third experiment investigated how the brain processes rhythmical information available in speech (sentences) and non-speech stimuli (artificial sounds, laughter). As a function of speech processing left perisylvian activation is detected. Artificial sounds primarily recruit bilateral areas deeply buried in the posterior part of the Sylvian fissure, whereas laughter corresponds solely to right cortical areas, in particular in auditory and (sensory-)motor cortices. Taken together, the results demonstrate how fMRI helps to gather new insight about the neurocognition of speech by identifying distinct cerebral regions subserving in distinct functions of speech.

11/12/01 Modelling the firing patterns of Mitral cells

Rebecca Smith

I will begin with a brief summary of my in vivo experiments and the data I collected for the benefit of those who haven't heard any of my talks before. I will then go on to talk in detail about Andrew Davison's model of a Mitral Cell which I am adapting so that it can reproduce my experimental results.

04/12/01 From SUPERstition to UNDERstanding: what's in between?

Nicola De Pisapia

In this talk we will discuss some experiments and models in animal conditioning that computational theories of the human Prefrontal Cortex should take into account. These studies (in particular Balleine and Dickinson 1998) show that Skinnerian stimulus-response habit mechanisms, producing the so-called superstitious behaviour, cannot fully explain goal directed behavior as experimented in mammals. In particular, one fundamental missing part is the capacity of mammals to learn the contingency (causal relationship) between stimuli and actions, or, in other words, to know and evaluate the consequences (states of the environment) of their actions before they actually execute them. Stimulus-response habits can instead only detect temporal contiguity. We will see this with a Reinforcement learning simulation. The capacity to "understand" the causal relationship presupposes the use of an "internal model of the world" that predicts the temporal evolution of the environment. Lesion studies in the Dorsolateral Prefrontal Cortex of rats suggest that this is the part of the brain mainly involved in contingency learning. We will re-interpret the role of the human Dorsolateral Prefrontal cortex in Planning as that of a builder of such internal models.

20/11/01 Flytrap - a database system for targeted gene expression
studies in the Drosophila brain

Douglas Armstrong

The application of transgenic technologies to the structure and function of the nervous system has largely been pioneered in the fruit fly Drosophila melanogaster. The diverse range of research interests than span the Drosophila neuroscience community highlights the inherent complexity of the nervous system. A new generation of collaborative tools built using adaptive principals are required to fully exploit current and future transgenic technologies. This presentation will describe the development of an established research resource ( and describe plans to develop it into a collaborative tool that can support synergistic research collaborations.

06/11/01 Extending the Linsker network to model feedback in visual processing

Tim Hely

In this talk I will discuss an extension of the Linsker feedforward neural network (Linsker, PNAS 83: 7508-7512, 8390-8394; 1986) which I am using to investigate the development and functional role of cortico-thalamic feedback connections. Despite the fact that many more connections project back from the visual cortex to the thalamus than feed forward from the retina, the role and significance of this massive pathway is still unclear. The Linsker network was one of the first models to show how the receptive field properties of cells in the early visual pathway could develop without "structured" input - in particular those of ON/OFF cells in the retina and simple orientation cells in V1. I will discuss the results of applying the same techniques to the development of the receptive fields of feedback connections. These results agree with the biological data with one important difference. I will also discuss preliminary (and rather inconclusive) results from the main goal of this work which is to investigate the functional role of multiple feedforward/feedback filtering passes in visual processing. The majority of this project was carried out at the Santa Fe Institute and is very much ongoing work in progress.

30/10/01 Evolving NeuroML Models

Paul Rogister

NeuroML is a project aimed to create a language that can describe every aspect of a neuroscience model. Such language would have many benefits, the most wanted being to allow researchers to easily write, share and exchange models reusable on every simulators. But besides those targeted benefits of NeuroML, this talk will discuss one of the unexpected effect of NeuroML : opening the way to neuroscience model evolution. Model evolution, largely used in Artificial Neural Network, can greatly improve model definition, but are the methods used with Artificial Neural Network applicable to more complex neuroscience models? We will try to show here that some of the difficulties in evolving models can be solve by using NeuroML as a genes coding to be used in an evolutionnary algorithm, and that even if all obstacles are far to be levelled, this could lead to an interesting disciplines : improving neuroscience research by applying evolution on its models.

23/10/01 Error Correction and Cryptography

David Barber

In a noisy environment, the reliable transmission of information can be secured by using a redundant representation of data. Such scenarios occur in both biological and man-made systems. I will describe some of the general basics behind the pervasive concept of error-correction, and also try to impart some of the wild excitement that has gripped the coding community for the last five years since the re-discovery of codes with extremely impressive performance. The flip side of error-correction, cryptography, will be explained and, in particular, the direct application of graphical models as a public-key cryptosystem.

16/10/01 Task-switching and compatibility between
stimulus attribute and response mode

Rayna Azuma

We examine the effect of compatibility between stimulus attribute and response mod on the cost to performance of switching between tasks. The stimulus was a rectangle containing "LEFT", "RIGHT", or "XXXX" on its left, right, or in the middle. Subjects switched, every second trial, between responding to the word (LEFT/RIGHT) and indicating which side (left/right) the string was on. Mean RT and interference effects confirmed the intuition that saying "left" to LEFT is more compatible than pressing a left key, while responding to the side with a keypress is more compatible than naming it. We show that (a) switching between the two compatible S-R mappings was less costly than switching between two less compatible mappings; (b) switching between more and less compatible mappings to the same response mode yielded symmetrical switch costs (c) compatibility influenced the costs of switching to the side task from , not the cost of switching from the side task to an unrelated task; (d) results (b) and (c ) were not appreciably changed by using only neutral stimuli (XXXX on one side, or a centred LEFT/RIGHT). These results seem problematic for Allport, Styles and Hsieh's (1994) claim that the dominant task is harder to switch to, and for their Task Set Inertia theory. We favour an account of switch cost as time consumed by control operations of enabling and disabling S-R mappings and pathways.

09/10/01 Implications of more
realistic temporal plasticity
rules in more realistic neurons

David Sterratt

Members of ANC with keen memories will have noticed that the title for this talk is remarkably similar to that of the last one I gave. The major difference is that this talk deals with more realistic temporal plasticity rules as well as more realistic neurons.

In the temporal plasticity rule studied by Song & al. (2000), the amount of potentiation or depression at a synapse is determined by the relative timing of spikes in the pre- and postsynaptic neurons associated with the synapse. Depression is stronger than potentiation and the weight strengths are strictly limited. In a situation where many randomly-firing pre-synaptic neurons synapse onto a single postsynaptic neuron, this rule leads to a bimodal weight distribution.

Van Rossum & al. (2000) study a more realistic plasticity rule in which the amount of potentiation or depression depends continuously on the current strength of the synapse as well as the relative timing of pre- and postsynaptic spikes. This model leads to a more realistic unimodal weight distribution in the same situation as above.

What happens when the neuron model is made more realistic by spreading the inputs over the dendritic tree? Assuming that the weight change depends only on information available at the synapse, the effective plasticity rule will be different at different points on the dendritic tree. Similarly, the distribution of pre-synaptic spikes relative to postsynaptic spikes will be vary divisive and subtractive threshold mechanism is possible for an associative memory network.

15/05/01 A Biased Pop-Philosophical History of Probability and Statistics

Amos Storkey

This talk will look at how the shift in epistemology towards foundationalism, combined with similar pressures from the social and political sciences shaped the development of both frequentist and Bayesian statistical approaches. Though the founding frequentist statisticians such as Venn were explicitly empiricist, foundationalist influences can also be seen in the work of both J. M. Keynes and Jeffreys. However Keynes explicit subjectivism combined with his rejection of the principle of indifference mean his work never delivers the test for knowledge that he aims for. Jeffreys use of the principle of indifference allows him to pursue objectivity further.

When criticisms of foundationalism, empiricism and logical positivism are encountered from a number of directions, especially postmodern writers, does Keynes subjective view of probabilities stand firmer despite his methodological inadequacies? Once the prior as degree of subjective belief is accepted, there is no room for foundational methodology. I argue that it is not until the foundational pressures are rejected that Bayesian inference actually makes sense. Thus a question can be posed... Is Keynes' forthright subjectivism a defining influence on the current direction of Bayesian philosophy as much as Jeffreys' is for defining Bayesian methodology?

08/05/01 Implications of temporal learning rules in more realistic neurons

David Sterratt

I will start by presenting the experimental evidence for a temporal learning rule at synapses between spiking neurons. After briefly discussing possible mechanisms for the rule and functions it might serve, I will present Song & al's (2000) model. The model comprises a single integrate-and-fire neuron with many excitatory and inhibitory random spiking function units in modelling hemispheric asymmetries, and illustrate in an implemented model how different receptive field size can account for these asymmetries in neglect.


2/5/00 Variable memory Markovian source modeling
and application to protein family characterization

Gill Bejerano

A general learning scheme for modeling discrete sources will be presented. This scheme is appropriate for sources exhibiting short memory behaviour, meaning that a typical symbol in a sequence generated by such a source can be fairly accurately estimated based on a relatively short segment of the sequence immediately preceding it. Many natural sources are known to display this property. An algorithm implementing this learning scheme in optimal time and space requirements will be outlined. Finally, i will present a successful application of the method to the problem of protein family characterization.

Based on joint works with Golan Yona from Stanford and Alberto Apostolico from the universities of Purdue and Padova.

18/4/00 Advances in Kernel Methods

Chris Williams

In this talk I shall discuss 2 new developments in kernel methods:

1) On a Connection between Kernel PCA and Metric Multidimensional Scaling (submitted to Machine Learning)

2) The Effect of the Input Density Distribution on Kernel-based Classifiers (accepted to ICML-2000)

Abstract 1)

In this note we show that the kernel PCA algorithm of Schölkopf \etal\ (1998) can be viewed as a form of metric multidimensional scaling (MDS) when the kernel function $k(\bfx, \bfy)$ is isotropic, i.e.\ it depends only on $||\bfx - \bfy||$. This leads to a metric MDS algorithm where the desired configuration of points is found via the solution of an eigenproblem rather than through the iterative optimization of the stress objective function. The question of kernel choice is also discussed.

Abstract 2)

The eigenfunction expansion of a kernel function $K(\bfx, \bfy)$ as used in support vector machines or Gaussian process predictors is studied when the input data is drawn from a distribution $p(\bfx)$. In this case it is shown that the eigenfunctions $\{ \phi_i \}$ obey the equation $\int K(\bfx, \bfy) p(\bfx) \phi_i(\bfx) d\bfx = \lambda_i \phi_i(\bfy)$. This has a number of consequences including (i) the eigenvalues/vectors of the $n \times n$ Gram matrix $K$ obtained by evaluating the kernel at all pairs of training points $K(\bfx_i, \bfx_j)$ converge to the eigenvalues and eigenfunctions of the integral equation above as $n \rightarrow \infty$ and (ii) the dependence of the eigenfunctions on $p(\bfx)$ may be useful for the class-discrimination task. We show that on a number of datasets using the RBF kernel the eigenvalue spectrum of the Gram matrix decays rapidly, and discuss how this property might be used to speed up kernel-based predictors.

4/4/00 Topographic mapping, human language and the mental lexicon

Richard Shillcock

Spoken word access is typically investigated by means of on-line experimentation, yet the lexicon itself contains a wealth of structure from which we can make inferences about processing. I review studies of the statistical structure of the English lexicon and conclude that the lexicon is shot through with partial but statistically significant correspondences between form and meaning/function, reflecting the brain's predisposition for creating topographic mappings. I argue that this structure is the result of the language adapting to the cognitive constraints of the listener/speaker. The model of the language processor that emerges both from first principles and from new data is one in which topographic mappings of different types of information interact, with a tendency towards isomorphism between those mappings. The data are consistent with a genuine flow of information from "higher" to "lower" levels of representation and processing.

18/1/00 fMRI with Arterial Spin Labeling

Enrico Simonotto

Arterial Spin Labeling is an MRI technique that allows one to make direct, and in some cases absolute, measurements of regional cerebral blood flow. Recently, it has been demonstrated that ASL can be used to locate functional activation in the brain. I will briefly introduce ASL, describe how it can be used for fMRI and compare ASL to the more commonly used BOLD contrast.

My talk is based on Chap. 4,5,6 and 17 of:

Functional MRI, Moonen, C.T.W. and Bandettini, P.A. (Des), Springer (1999)

Connectionist models of infant language learning

Nigel Goddard

I will review a recent study of infant language learning (grammar) and will compare two connectionist models which claim to account for the data. The first is a model by Seidenberg and Elman which uses the Elman recurrent network structure. The second is a model by Shastri which combines his binding-with-synchrony technique with Watrous's method for

16/11/99 Perception as Bayesian Inference

Chris Williams

A presentation of "Perception as Bayesian Inference" (eds Knill and Richards), focusing on chapter 1 (Mumford: Pattern Theory, a unifying perspective) and chapter 12 (Barlow; Banishing the homunculus).

Plus a quickie discussion on graph theory relating to finding hubs and authorities in a web search. learning in recurrent networks with delays.

9/11/99 More on split foveas, split brains and word recognition

Richard Shillcock

I will present the final state of the research that Padraic Monaghan and I have carried out looking at the effects of coordinating information across the two halves of a split processor. The goal is to help to reconceptualise visual word recognition along lines that reflect the basic anatomy of the human visual system. The earlier story we had has now become a little more complex ...

I will talk a little about what psychologists might hope to learn from such modelling.

I will also talk about work in progress with Lindsay Payne and Padraic Monaghan looking at some perplexing data from deep dyslexia concerning the role of prosody in visual word recognition.

5/10/99 Products of Experts

Professor Geoffrey Hinton (Gatsby Computational Neuroscience Unit, UCL, London)

It is possible to combine multiple probabilistic models of the same data by multiplying the probabilities together and then renormalizing. This is a very efficient way to model high-dimensional data which simultaneously satisfies many different low-dimensional constraints. Each individual expert model can focus on giving high probability to data vectors that satisfy just one of the constraints. Data vectors that satisfy this one constraint but violate other constraints will be ruled out by their low probability under the other expert models. Training a product of models appears difficult because, in addition to maximizing the probabilities that the individual models assign to the observed data, it is necessary to make the models disagree on unobserved regions of the data space: It is fine for one model to assign a high probability to an unobserved region as long as some other model assigns it a very low probability. Fortunately, if the individual models are tractable there is a fairly efficient way to train a product of models. This training algorithm suggests a biologically plausible way of learning neural population codes.

25/8/99 Chaos and Noise in the Brain (and what to do about them)

Mark L. Spano (US Navy, Carderock Laboratory)

The brain is a highly connected system of nonlinear elements. So it should not be surprising that is can exhibit the same types of behavior as other complex nonlinear systems - namely chaos and stochastic resonance.

The first part of the talk will focus on the manifestation of chaos during epileptiform activity in a rat brain hippocampal preparation. Attempts at controlling the chaos will be discussed.

The second part of the talk will look at how the same preparation can exhibit stochastic resonance. This ability to enhance signals buried in noise may shed some light on the brain's ability to process information.