Learning Dynamics, and Dynamics of Learning.
Event Details
- Type
- Center for Studies in Physics and Biology Seminars
- Speaker(s)
-
Ilya Nemenman, Ph.D., Samuel Candler Dobbs Professor, Emory University
- Speaker bio(s)
-
The Rescorla–Wagner model of associative learning is a staple of animal learning theory. I will start with the model's strong spots, but then show, by analyzing two experiments — in songbirds and in worms — that it cannot give a quantitative account of the dynamics of learning in these animals. The worm needs multiple associative channels of differing sign, magnitude, and learning rate to explain spontaneous recovery and other features of its behavior. The bird is better described by Bayesian filtering with non-Gaussian error channels. And both animals generalize what they learn to different contexts. I will show how modified related models explain the data much better, and what this implies for how learning is implemented in real nervous systems and should be implemented in artificial ones. Switching gears, I will then walk you through our lab's attempts to learn biological dynamics directly from data, with machine learning that keeps the inferred models interpretable — from the nociceptive response of C. elegans to our recent attempts to write a fairly generic dynamics inference agent, which I hope some of you may want to use.
- Open to
- Public
- Phone
- (212) 327-8636
- Sponsor
-
Melanie Lee
(212) 327-8636
leem@rockefeller.edu