53 min
Seminar, Hosted by Center for Brains, Minds and Machines (CBMM), MIT
Invariance and equivariance in brains and machines
UC Berkeley
Recording
Abstract
Building machines that perceive and act in the world as humans and other animals do has been a focus of AI research for over half a century, while neuroscience has sought a mechanistic understanding of the brain processes underlying perception and action over the same period. These parallel efforts could inform one another, yet recent advances in deep learning and transformers have mostly not translated into new neuroscientific insight, and AI, beyond loose inspiration from neuroscience, has largely pursued its own course that now deviates strongly from the brain.
This talk proposes an approach to building both invariant and equivariant representations in vision that is rooted in observations of animal behavior and informed by neurobiological mechanisms (recurrence, dendritic nonlinearities, phase coding) and mathematical principles (group theory, residue numbers). The result is a neural circuit for factorization that can learn about shapes and their transformations from image data, and a model of the grid-cell system based on high-dimensional encodings of residue numbers. These models provide efficient solutions to long-studied problems, are well suited to implementation in neuromorphic hardware, and offer a basis for hypotheses about visual cortex and entorhinal cortex.