Cosyne 2007 Workshops
February 27, 2007
The Canyons, Utah
Speaker Name
Peter Latham, Lecturer, Gatsby Computational Neuroscience Unit, University College London
Talk Title
Unsupervised learning, correlations, and error correcting codes: how are they related, and what can neural data tell us about them?
Talk Abstract
There are two main strategies the brain could use for representing information: analog and digital. In the former, continuous variables can take on a continuum of values. Color, for example, could take on any value in some three-dimensional RGB space. In the latter, variables are discretized, so color could take on only a finite number of values. While discretization might seem artificial for variables that are naturally continuous, this approach comes with a huge advantage: the brain can make use of error correcting algorithms, and thus transmit information with arbitrarily small loss.
Determining which strategy the brain uses is important for understanding how it represents and manipulates information. It's also important because it should provide insight into a deeper question: how does the brain makes sense of the world? Here we show that the analog versus digital strategies can be distinguished by looking at the scaling of total entropy versus the number of observed neurons in a population, denoted N. Specifically, for analog coding schemes, the entropy scales as aN+b+c*log(N), whereas for digital schemes it scales simply as aN+b. Thus, determining coding strategies comes down to computing entropy versus N and looking for a log(N) correction.
