DEC Colloquium

Principles of Neural Design

Peter Sterling (UPenn)
Practical information
15 September 2015

The supercomputer that beat Garry Kasparov at chess fills a room and draws megaWatts; whereas our brain occupies 1.3 liters and draws only 12 watts. Therefore, brain design must reflect intense constraints on space and energy. One key is to compute with chemistry because diffusion is fast and cheap over short distances (micrometers). To transmit beyond that requires electrical signals, which are expensive, costs rising disproportionately with information rate. Consequently neural designs try to hold the steep part of the rate vs cost curve by sending only information that is needed, by sending it slowly, and by shortening the wires. To conserve space, the brain saves new information only for as long as it will be needed and couples learning to forgetting. To capture the overall design ten principles suffice.