Bodian Seminar: Bryan Tripp, PhD
Centre for Theoretical Neuroscience
University of Waterloo
“Toward ethologically realistic models of neural systems”
Computational models are helpful for understanding neural systems, due to their complexity and nonlinearity. Existing neurophysiological models are realistic in various dimensions, such as cell morphology and stimulus tuning. However, there is a relative lack of functionally realistic models, i.e. models that are capable of behavior that is relevant to the natural lives of animals, and thus the actual survival value of the brain. It will be hard to close this gap, but I will argue that the process can be accelerated with deep learning. Notably, existing deep networks perform a variety of ethologically relevant tasks in ways that overlap human behavior somewhat. However, this is largely accidental, as there has been little effort to use deep learning specifically to model neurophysiological circuits. I will discuss our ongoing work in this direction, including optimization of a deep architecture to approximate the architecture of the primate visual system, incorporation of a contour-‐integration model, and efforts to obtain realistic visual representations. I will also discuss our related work in robotics, to develop realistic functional tests of neurophysiological models.