Background
I started out as an engineer wanting to build robots inspired by how brains work, and I ended up as a neuroscientist attempting to understand how nervous systems process information, inspired and guided by engineering principles. I first learned about neural networks as a student at Stanford, through Bernie Widrow’s course on Adaptive Filters and Misha Pavel and Ilan Vardi’s Connectionist Models seminar (1986/87). I then found my way to Pentti Kanerva’s Sparse Distributed Memory (SDM) research group at NASA/Ames, where I worked for two years as a research assistant to develop vision applications of SDM. During this time I learned about Charlie Anderson and David Van Essen’s work on ‘shifter circuits,’ which eventually led to my doing a Ph.D. under their joint supervision as a student in the CNS program at Caltech in the early 1990’s. My thesis was on dynamic routing circuits, essentially a generalization of shifter circuits which could serve as a neural mechanism for forming position and size (and rotation) invariant representations in the visual cortex. Toward the end of my thesis work I learned about David Field’s work on models of sensory coding based on natural image statistics, which seemed like a promising way to learn feature representations at various stages of the cortical hierarchy. It remains on of my central goals to bring these two ideas – i.e., dynamic routing and feature learning – together in order to build a hierarchical model of the visual cortex. I began my academic career as an Assistant Professor in the Department of Psychology (and later, Neurobiology, Physiology and Behavior) and Center for Neuroscience at UC Davis. In 2002, together with Pentti Kanerva and Fritz Sommer, I helped Jeff Hawkins to launch the Redwood Neuroscience Institute (RNI). In 2005 RNI was incorporated into UC Berkeley’s Helen Wills Neuroscience Institute and renamed the Redwood Center for Theoretical Neuroscience, for which I currently serve as director.