Login
Computer Science @ Rochester
Wednesday, April 28, 2004
11:00 AM
CSB 209
Ph.D. Thesis Proposal
Jonathan Shaw
University of Rochester
Efficient Coding and Bayesian Inference
The brain deals with a vast amount of information coming in from the body's senses, and performs compuation on this input using spiking neurons. Understanding this computation requires an understanding of two things: how the spikes represent the input, and how the spikes can be combined to do useful work. There is currently no consensus on how either of these is done. One possible answer to the coding question is that the code used is the code which represents the input using the minimum number of spikes. This principle is called efficient coding, and, in addition to having a number of desirable properties, it makes predictions about the receptive fields of the neurons in the visual cortex which fit very well with those observed in the primate visual cortex. However, there are very few models which explain how efficient coding can actually be achived using spiking neurons, and these models have very strong and somewhat dubious assumptions behind them. I will talk about my own model which performs linear efficient coding but avoids making some of these assumptions, as well as the future extensions I plan to make to the model which will allow it to do nonlinear efficient coding.

Optimal computation in the brain is usually cast in the form of Bayesian inference. It has long been known that different computations can be made easy or made difficult depending on how the data to be used is represented. It therefore seems natural to work on Bayesian inference using data which is represented using efficient coding. However, there has been no work from the Bayesian inference field for using efficient coding, and there has been no work from the efficient coding field on performing Bayesian inference. I will argue that efficient coding automatically performs Bayesian inference. The focus of my research is therefore to demonstrate this link between efficient coding and Bayesian inference.