CS Seminar Series: Martin Falk

Learning to Learn in Physical Systems

Abstract:

Biological and molecular systems can learn surprisingly diverse computational tasks, but straightforward abstractions of learning with artificial neural networks often map poorly onto these noisy and decentralized substrates. In this talk, I will discuss a different approach for achieving computation and learning, one which leverages the natural dynamics of physical systems. I will show how plasticity and high-pass filtering can combine to produce a form of contrastive Hebbian learning; how related molecular processes can support Pavlovian conditioning and supervised classification in single cells; and how exposure to changing training tasks can produce physical systems that rapidly switch between otherwise incompatible functions. Together, these results suggest new frameworks for understanding how learning can arise from simple physical processes, and how even in simple physical systems, the ability to learn can be shaped and evolve over time.

Bio:

Martin Falk is an incoming Assistant Professor of Physics at the University of Massachusetts Amherst. He currently is at the University of Chicago, where he held a Schmidt AI in Science Fellow and worked with Arvind Murugan and William Irvine. Prior, he completed his Ph.D. in the MIT Physics Department as an NDSEG Fellow, supervised by Michael Brenner and Leonid Mirny. He received his B.A. in Math and Physics from the University of Pennsylvania.