Prof. David Schwab: Deconstructing Collective Neural Activity with Hidden Variables
January 21, 2015
Bausch and Lomb Lobby
Prof. David Schwab
Deconstructing Collective Neural Activity with Hidden Variables
Recently it has become possible to directly measure the simultaneous activity of large populations of neurons. Strikingly, the underlying probability distributions of the collective states of these ensembles often exhibit a feature known as Zipf’s law. That is, the frequency of a state is inversely proportional to its rank in a frequency table.
Translating this into the language of statistical physics reveals that these systems appear poised near a unique critical point, where the extensive parts of the entropy and energy are exactly equal. Here we present analytical arguments and numerical simulations showing that such behavior naturally arises in systems with unobserved random variables, such as a common input stimulus to a neural population, that affect the observed degrees of freedom. We then build on these results to motivate a class of models for collective neural activity based on so-called deep neural networks. After reviewing this class of models, we show that they can be viewed as implementing a form of real-space renormalization. We then fit this model to neural population data, demonstrating its utility to capture subtle changes in higher-order correlation structure.
University of Rochester
Physics and Astronomy Colloquium
Date: Wednesday, January 21, 2015
Tea: 3:30 pm, Bausch and Lomb Lobby
Talk: 3:45 pm, Bausch and Lomb 106