Computer Science Colloquia
Wednesday, October 22, 2014
William "Chip" Levy
Host: Worthy Martin
3:30 PM, Rice Hall, Rm. 130 (Auditorium)
Defining Neural Computation Through Optimization Results
Using the optimizing nature of natural selection and half a billion years of animal evolution, one can expect that microscopic neural processes are optimal. Just a few such optimization results are known but these results all support minimizing the energy cost of communication and computation. However, under pervasive, simplifying assumptions, a neuron, described by the signal transformation it performs, seems far, far from optimal >108 times the limit set by statistical thermodynamics. Here we re-analyze and re-interpret neuronal physiology, carefully distinguishing communication costs within a neuron from computational costs, and arrive at a new definition of microscopic neuronal computation. This distinction reveals that communication accounts for most of the extremely large costs (99.91%). As a result of this reanalysis, one aspect of computation is predicted to occur near the best possible Joules per bit limit set by physics.