Personal tools
You are here: Home Events Workshop: Philipp Rudiger and Krzysztof Geras, Chair: Botond Cseke

Workshop: Philipp Rudiger and Krzysztof Geras, Chair: Botond Cseke

— filed under:

  • ANC Workshop Talk
When May 14, 2013
from 11:00 AM to 12:00 PM
Where IF 4.31/4.33
Add event to calendar vCal

Philipp Rudiger

Unifying anatomical, psychophysical, and developmental circuit models of primary visual cortex

A variety of proposed circuits have been introduced for the primary visual cortex of mammals (V1), each explaining behavior under a narrow range of potential inputs and temporal scales. Electrophysiological evidence shows that V1 neuron responses differ markedly depending on contrast and distance [1]. Researchers studying long-term development generally assume Mexican-Hat lateral interactions, and psychophysical studies of surround modulation have primarily focused on long-range inhibition, leading to the term ''surround suppression''. Yet the actual anatomical substrate for long-range interaction appears to mediated by excitatory connections [1], and several proposed circuits have focused on long-range excitation's role in contour completion and pop-out [2].

Here we propose a model that explains how each of these sources of evidence is compatible with a single underlying circuit, which effectively reduces to Mexican-Hat connectivity for high-contrast inputs. The model is based on two anatomically and functionally distinct populations of inhibitory interneurons: basket cells, i.e. wide-arbor perisomatically targeting cells with untuned, low-latency suppression [3,4], and somatostatin-expressing cells, which are smaller, high-latency, and strongly tuned, potentially providing long-range orientation-specific inhibition through polysynaptic circuits [5].

This approach ensures the model is constrained not only by electrophysiological and psychophysical data but also the need for robust map development and anatomically realistic connectivity. The model allows for the development of strong lateral excitatory connections, which in combination with the different sources of surround suppression, allows the model to exhibit realistic low-contrast behaviors such as iso-orientation facilitation and contour completion, in addition to contrast-dependent size-tuning shifts and pop-out displayed by some of the earlier models. We analyze the roles of the two inhibitory populations in playing a permissive role for long-range excitation at low contrasts, and mediating surround suppression at higher contrasts. In doing so, we demonstrate, for the first time, a clear link between development of orientation selectivity and the underlying connectivity, with the perceptual consequences of this circuitry.

In future, this work will help us to complete our understanding of the V1 circuit by adding feedback and neuromodulatory mechanisms. Additionally, the results may be used to provide realistic connectivity patterns for large-scale spiking models, which often struggle to adequately constrain their connectivity.

[1] Hirsch & Gilbert J. Neurosci. (1991) 11:1800-9.

[2] Stemmler et al. Science (1995) 269:1877-80.

[3] Buzas et al. J. Comp. Neuro. (2001) 437:259-85.

[4] Lund et al. Cereb. Cort. (2003) 13:15-24.

[5] Adesnik et al. Nature (2012) 490:226-31.


Krysztof Geras

Multiple-source cross-validation

Cross-validation is an essential tool in machine learning and statistics. The typical procedure, in which data points are randomly assigned to one of the test sets, makes an implicit assumption that the data are exchangeable. A common case in which this does not hold is when the data come from multiple sources, in the sense used in transfer learning. In this case it is common to arrange the cross-validation procedure in a way that takes the source structure into account. Although common in practice, this procedure does not appear to have been theoretically analysed. We present new estimators of the variance of the cross-validation, both in the multiple-source setting and in the standard iid setting. These new estimators allow for much more accurate confidence intervals and hypothesis tests to compare algorithms.