From statistical mechanics to machine learning: effective models for neural activity

In the retina, the activity of ganglion cells, which feed information through the optic nerve to the rest of the brain, is all that our brain will ever know about the visual world. The interactions between many neurons are essential to processing visual information and a growing body of evidence suggests that the activity of populations of retinal ganglion cells cannot be understood from knowledge of the individual cells alone. Modelling the probability of which cells in a population will fire or remain silent at any moment in time is a difficult problem because of the exponentially many possible states that can arise, many of which we will never even observe in finite recordings of retinal activity. To model this activity, maximum entropy models have been proposed which provide probabilistic descriptions over all possible states but can be fitted using relatively few well-sampled statistics. Maximum entropy models have the appealing property of being the least biased explanation of the available information, in the sense that they maximise the information theoretic entropy. We investigate this use of maximum entropy models and examine the population sizes and constraints that they require in order to learn nontrivial insights from finite data. Going beyond maximum entropy models, we investigate autoencoders, which provide computationally efficient means of simplifying the activity of retinal ganglion cells.

Grantor

Department of Mathematics and Applied Mathematics

Year dc:date.issued

2022

Author dc:creator

• Schonfeldt , Abram

Advisors dc:contributor.advisor

• Rohwer, Christian
• Shock, Jonathan

• applied mathematics

Handle dc:identifier.uri

http://hdl.handle.net/11427/37848

OAI identifier oai:identifier

oai:open.uct.ac.za:11427/37848