Relational Architectures

Abstract:
Despite many decades of intensive research, the basic principles of cortical computation remain a mystery. Even the format of the data continues to be debated (e.g. average population activity vs. spike timing). Visual areas are the most studied and best understood in the cortex, but even in V1 we do not understand the operation of the cortical circuitry. However, we do know a lot about the types of responses that are seen in individual neurons. I will propose that the types of responses we see correspond closely to signals in relational computation, a style of computation that allows inferences to flow in any direction through a network. (Belief propagation in factor graphs is an example of this style of computation.) I will argue that relational computation is a much more reasonable starting point for understanding cortical computation than the standard computer-inspired cpu/memory analogy. We will consider how this point of view can inform us on such diverse topics as population code structure, signal restoration, or the location of information in the cortex.

back