AASS Seminar - How structure in neural representations contributes to robust and efficient inference algorithms

The research centre AASS arranges a seminar with Christopher Kymn, UC Berkeley.

Remote attendance: https://oru-se.zoom.us/j/65096623293

Abstract

How can neuroscience inform algorithm development in computer science, and vice versa? In this talk, I will present two recent lines of effort by myself and collaborators on this question. First, I will introduce an algorithm for efficiently factoring visual scenes into their respective objects and locations. The framework extends previous algorithms based on neuroscience, including sparse coding and dynamic routing. Second, I will introduce a mathematical model of the hippocampus and entorhinal cortex possessing a large memory capacity and capable of performing robust path integration. The model makes useful predictions for experimental neuroscience, but it can also lead to useful performance on optimization algorithms. Both systems rely on the properties of random high-dimensional vectors, and on binding mechanisms proposed in cognitive science. These results suggest that binding operations are a useful primitive in neural computation for both brains and machines.

Short bio

Christopher Kymn is a final year Ph.D. student in the Redwood Center for Theoretical Neuroscience at University of California, Berkeley. He received a M.Sc. in neurosciences from LMU Munich and an A.B. in computer science and philosophy from Dartmouth College. His current research interests center around theoretical neuroscience and machine learning, using techniques from hyperdimensional computing, sparse coding, and attractor neural networks.