home · mobile · calendar · colloquia · 2006-2007 · 

Colloquium - Anderson

A Comparison of Neurobiological and Digital Computation
Washington University in St. Louis

There is overwhelming evidence that neuronal systems represent analog variables with highly redundant population codes, a simple idea with a long history in neuroscience. Indeed, this principle underlies all neural prosthesis research where a few tens of electrodes interact with many hundreds to thousands of neurons. Starting from this premise, my students and I have developed a computational framework that shows how large networks of spiking neurons can store and transform analog signals for sensory processing, motor control, and statistical inference. The resulting computational systems differ from traditional artificial neural networks that are focused on the highly nonlinear properties of individual neurons, and are more in line with modern Bayesian systems. The brain is more like an analog computer than a digital one; more like a Bayesian inference machine than a symbolic one.

This talk will touch on these and other topics, contrasting the way digital and neurobiological systems handle common implementation issues such as the storage and transformation of variables, routing of information, virtual processing, etc. Some speculations will be made on a generalized concept of pointers in which something like symbolic processing might be implemented in the brain. However, the emphasis is on low level processing issues and not on how the brain is organized to carry out high level cognitive processes.

Hosted by Michael Mozer.

Department of Computer Science
University of Colorado Boulder
Boulder, CO 80309-0430 USA
May 5, 2012 (14:13)