Personal tools
You are here: Home Events Past Events Past Events (2009)

Past Events (2009)

17/11/09  11:00 - 12:00

ANC Seminar: Tony Prescott (Host Jim Bednar)

Understanding brain architecture through active touch sensing in animals and robots 

The systems approach in the brain sciences has demonstrated that there is no straightforward decomposition of the brain into modules, or even a simple means to separate the brain from the body (in control terms), or the body from the environment. So how should we proceed to understand the relationship between brain and behaviour? Our approach has been to investigate a complete sensorimotor loop, specifically, the guidance of exploratory behaviour by tactile sensing signals. We have focused on the rat whisker (vibrissal) system as a model. The neurobiology of this system indicates multiple layers of control, that can be loosely mapped to the different levels of the neuraxis, and that exhibit both some redundancy and some modularity. Neuroethological experiments show a tight coupling between sensory signals and active control of the movement and positioning of the sensors. Electrophysiological and modelling studies suggest a system that is capable of rapidly extracting relevant affordances for action, rather than constructing complex internal representations of the external world. These ideas will be illustrated with examples from our research on active touch sensing in animals and in biomimetic robots.


27/10/09  11:00 - 12:00

ANC Seminar: Andrew Golightly (Host Amos Storkey)

Andrew Golightly (Newcastle University) 

Bayesian Inference for Hybrid Discrete-Continuous Systems Biology Models

We consider the problem of efficiently inferring the parameters in gene regulatory networks.  Whilst it is possible to work with a discrete stochastic model for inference, computational cost can be prohibitive for networks of realistic size and complexity. By treating the numbers of molecules of biochemical speciesas continuous, a diffusion approximation can be used, and whilst this approach has been shown to work well for some networks, ignoring discreteness of low copy number species is unsatisfactory. In this paper, we consider a hybrid inference method, treating low copy number species as discrete and the remaining species numbers as continuous. The methodology uses a recently proposed hybridsimulation scheme inside a particle filter. We apply the scheme to a simple application and compare the output with a scheme for performing inference for the underlying discrete stochastic model. 


29/09/09  14:00 - 15:00

ANC Seminar: Paul Fearnhead - This is a joint seminar with the Statistics Group, School of Mathematics.

Paul Fearnhead (University of Lancaster)

Efficient Bayesian analysis of multiple changepoint models 

We describe an efficient algorithm for Bayesian analysis of multiple changepoint models. In many scenarios it enables iid samples from the posterior distribution. Approximate versions (which introduce negligble error) have a computational cost that is linear in the number of observations -- and thus can be applied to large data sets (such as arise in modern bioinformatic applications). 

The method is demonstrated on applications that range from inference about the divergence of Salmonella Typhi and Paratyphi A, to inference about the Isochore structure of the human genome. 


15/09/09  11:00 - 12:00

ANC Seminar: Peter Kind

Glutamate-Dependent Cortical Development and Cognitive Impairment


07/07/09  11:00 - 12:00

ANC Seminar: Sophie Deneve

Contextual effects in visual processing as optimal probabilistic inference. 

Stimuli outside the classical receptive field (cRF) of a visual neuron can nevertheless strongly modulate the cell's response to stimuli inside the cRF. The extent and influence of the cRFs and ncRFs, and the amplitude of the response also varies as a function of contrast, time, stimuli in the surround and previous stimuli. In particular, these observations hint at a coarse to fine coding continuum, a switch from integration with the context  to competition with the context (i.e. saliency, adaptation) as more sensory information becomes available. 

Here we show that cRF and ncRF spatial properties, their temporal dynamics , adaptive properties and contextual dependencies can emerge naturally from detecting objects in visual movies. This approach draws a parallel between the  spatio-temporal statistics of the visual input and the plasticity and dynamics of a network of adapting integrate and fire neurons. It characterizes visual cells, not by  their receptive field (i.e. spatio-temporal filters) but by their "causal fields" (i.e. the predicted impact of the feature they represent on the visual input). 

This approach provides several new insights about the nature of visual  processing. In particular, divisive inhibition (as opposed to subtractive inhibition), and spike-based computations emerge as an essential components of sensory processing and perception in a dynamical, ambiguous world. 


23/06/09  11:00 - 12:00

ANC Seminar: Geoff Hinton

Location: Lecture Theatre 1, Appleton Tower 

Geoffrey Hinton
Canadian Institute for Advanced Research
& University of Toronto 

A quick way to learn a mixture of exponentially many linear models. 

Mixtures of linear models can be used to model data that lies on or near a smooth non-linear manifold.
A proper Bayesian treatment can be applied to toy data to determine the number of models in the mixture and the dimensionality of each linear model but this neurally uninspired approach completely misses the main problem: Real data with many degrees of freedom in the manifold requires a mixture with an exponential number of components. It is quite easy to fit mixtures of 21000 linear models by using a few tricks: First, each linear model selects from a pool of shared factors using the selection rule that factors with negative values are ignored. Second, undirected linear models are used to simplify inference and the models are trained by matching pairwise statistics. Third, Poisson noise is used to implement L1 regularization of the activities of the factors. The factors are then threshold linear neurons with Poisson noise and their positive integer activities are very sparse. Preliminary results suggest that these exponentially large mixtures work very well as modules for greedy, layer-by-layer learning of deep networks. Even with one eye closed, they outperform Support Vector machines for recognizing 3-D images of objects from the NORB database. 


11/06/09  15:30 - 16:30

ANC Seminar: Lyle J. Graham

Lyle J. Graham
Neurophysiology of Visual Computation Laboratory
Laboratory of Neurophysics and Physiology, CNRS UMR8119
Université Paris Descartes, Paris, France 

Exploring the Functional Roles of Membrane Conductances with Experiments and Models – In Vivo Dynamic Clamp Study of Shunting Inhibition and the BK Current in Visual Cortex 

The essential dynamics of neurons underlying synaptic integration and spike generation are understood to define these cells' unique computational role within the neural net. Yet, in the context of the intact system and physiological function the quantitative and qualitative identification of specific biophysical mechanisms is still at an early stage. In this talk I will review our recent work in addressing these questions using in vivo dynamic-clamp recordings that are strongly tied to cellular and membrane models. In particular, we are characterizing the influence of shunting synaptic inhibition and the BK potassium current on the neuron's transfer function, probed with both artificial and functional, visual stimuli. In comparison to previous experimental and theoretical studies, we find that realistic shunting inhibition has a significant divisive effect on firing gain, as well as important effects on threshold and saturation. Shunting inhibition also has a non-linear effect on visual responses, reducing response amplitude but as well tightening response timing. We confirm predictions that the BK current facilitates spike firing, despite being a hyperpolarizing current. This effect is demonstrated by an increase in the gain of the f/I curve and of visual responses. 


02/06/09  11:00 - 12:00

ANC Seminar: Robert Jacobs

Department of Brain & Cognitive Sciences
University of Rochester 

Is Human Learning Optimal? 

We address the question "Is human learning optimal?" in two research projects. In the first project, human subjects performed a perceptual matching task in which they attempted to convert the shape of a comparison object into the shape of a target object. To do this efficiently required knowledge of the causal relations among a set of underlying hidden or latent variables. Our results indicate that subjects did achieve near-optimal performance levels, and they did acquire good knowledge of the causal relations. In the second project, subjects performed a visual pattern discrimination task in which patterns were linear combinations of a set of arbitrary "basis features". Associated with each feature was a noise parameter. Features with small noise values were reliable information sources for the discrimination task, whereas features with large noise values were unreliable sources. Subjects learned to combine the information from the features in a near-optimal manner. Both projects suggest that human learning is both optimal and non-optimal. It is optimal in the sense that it leads to near-optimal performances. It is non-optimal in the sense that learning requires many more training trials than required by an ideal learner. 



22/05/09  15:30 - 17:00

ANC Seminar: Leon Glass

Leon Glass 

Professor of Physiology
Department of Physiology
McGill University 

Predicting Perception of the Wagon Wheel Illusion 

Stroboscopic illumination of a rapidly rotating disk with radial spokes leads to a range of different stationary and moving images as the angular rotation frequency of the disk and the strobe frequency are varied. We compare predictions from the standard correlation model of motion perception with a model based on phase locking observed during periodic stimulation of an integrate-and-fire nonlinear oscillator. The close agreement between theoretical predictions and experimental observations suggests the possibility that periodic forcing of nonlinear neural oscillations may play a role in motion perception. 

Biographical note from David Willshaw: Leon Glass had a post doctoral stay here in Edinburgh in one of the departments which were the forerunners of the School of Informatics. He is renowned for his work on various complex systems including: neural networks, the theory of visual perception (the Glass effect), and complex rhythms in cardiac and respiratory systems (the Mackey-Glass equation). 


05/05/09  11:00 - 12:00

ANC Seminar: John Winn

Probabilistic models for understanding images. 

Getting a computer to understand an image is challenging due to the numerous sources of variability that influence the imaging process. The pixels of a typical photograph will depend on the scene type and geometry, the number, shape and appearance of objects present in the scene, their 3D positions and orientations, as well as effects such as occlusion, shading and shadows. The good news is that research into physics and computer graphics has given us a detailed understanding of how these variables affect the resulting image. This understanding can help us to build the right prior knowledge into our probabilistic models of images. In theory, building a model containing all of this knowledge would solve the image understanding problem. In practice, such a model would be intractable for current inference methods. The open challenge for machine learning and machine vision researchers is to create a model which captures the imaging process as accurately as possible, whilst remaining tractable for accurate inference. To illustrate this challenge, I will show how different aspects of the imaging process can be incorporated into models for object detection and segmentation, and discuss techniques for making inference tractable in such models. 



28/04/09  11:00 - 12:00

ANC Seminar: David Searls

Omic Empiricism 

The rise of omics signals a shift from hypothesis-driven to data-driven, "bottom-up" research, while systems biology stresses modeling and claims to move beyond reductionism. These trends are at opposite poles of a long-standing duality between empiricism and rationalism. Twenty-first century biology might benefit from a consideration of how philosophers have approached and attempted to reconcile this duality. 

BIO: David B. Searls is an independent consultant, until recently the senior vice president of bioinformatics at GlaxoSmithKline Pharmaceuticals.  He holds an adjunct appointment at the University of Pennsylvania, where he was formerly research associate professor of genetics.  He holds degrees from MIT, Johns Hopkins, and Penn. 


07/04/09  11:00 - 12:00

ANC Seminar: Bjoern Brembs

Learning by doing: the neurogenetics of multiple learning systems in Drosophila.

To make good decisions, animals and humans learn from the consequences of their earlier choices to guide later decisions. For instance, as toddlers we learn to say “please”, later, we learn how to behave at cocktail parties. Chimpanzees learn how to use a stick for termites. New-Caledonian Crows learn how to bend a wire for a food reward. Honeybees learn at what time of day they have to visit which flower patch. Such learning situations often consist of an early, exploratory phase and a later phase in which the behavior is reliably produced to exploit a resource. Using the genetic tools available in the fruit fly Drosophila melanogaster, we discovered that such complex tasks are accomplished by processing them with two separate, but interacting learning systems. The first, dominant system stores any relationships between stimuli in our environment. The second, anatomically and genetically distinct system is subordinate and directly modifies behavioral circuits. Their hierarchical interactions ensure that the information each animal acquires about its environment remains flexible for use in different circumstances (generalization). Prolonged exposure to a given situation reduces this flexibility in favor of a more efficient, stereotyped behavior (habit formation). In the fruit fly Drosophila, the molecular mechanisms of these two learning systems are starting to unravel. A prominent neuropil, the mushroom-bodies are mediating some of the interactions between these systems and regulate the balance between flexible exploration and efficient exploitation. 



17/03/09  11:00 - 12:00

ANC Seminar: Professor Geoff Goodhill

Professor Geoff Goodhill
Queensland Brain Institute, Brisbane 

Wiring the brain: modelling axon guidance and visual map development 

For nervous systems to function correctly they have to be wired up correctly. Through a combination of experiments and theoretical modelling we are attempting to uncover some of the computational rules that determine how appropriate patterns of wiring form during neural development. In this talk I will discuss some of our recent work in this regard, firstly on a Bayesian model of how axons detect molecular gradients, and secondly on how the statistics of visual scenes influence the structure of maps in the visual cortex. 


03/02/09  11:00 - 12:00

Network Diseases: Linking Developmental Changes in Brain Connectivity with Neural Dynamics

Marcus Kaiser
School of Computing Science
Newcastle University 

Brain connectivity, also called the connectome, has been a recent field of research identifying organizing principles of biological neural networks. The analysis of network structure and dynamics is part of the larger field of network science. This research has shown that neural systems exhibit a modular organisation at several hierarchical levels: from individual columns to clusters of cortical regions such as the visual cortex (Hilgetag & Kaiser, Neuroinformatics, 2004). In addition, neural systems show a distinct spatial organisation (Kaiser & Hilgetag, PLoS Comp Biol, 2006) with both reduced wiring as well as fast processing through a low number of intermediate steps. I will present a new model for the spatial development of neural networks (Kaiser et al. Cerebral Cortex, in press) and discuss how developmental factors can change the network organisation. Finally, I will look at how this structure is linked to neural dynamics such as oscillations or the activity spreading. A particular application of this research is understanding the developmental and topological factors that lead to
seizure spreading in epilepsy patients.