2nd MIAI Distinguished Lecture on July 12, 2022 - 3 PM CET

on the July 12, 2022

We are pleased to share with you the second MIAI Distinguished Lecture on July 12th, with Prof. Aaron Courville (LISA Laboratory, University of Montreal) who will give a distinguished lecture.

Parameter resets in supervised and reinforcement learning

In this talk I discuss recent work in my group exploring a curious finding: when training neural networks, periodically resetting some or all parameters can be helpful in promoting better solutions. begin with a discussion of our findings in supervised learning where we relate this strategy to Iterated Learning, a method for promoting compositionality in emergent languages. We then show how parameter resets appear to offset a common flaw of deep reinforcement learning (RL) algorithms: a tendency to rely on early interactions and ignore useful evidence encountered later. We apply this reset mechanism to algorithms in both discrete (Atari 100k) and continuous action (DeepMind Control Suite) domains and observe consistently improving performance. I conclude with some recent finding and speculation about the underlying causes behind the observed effects of parameter resets.
MIAI Seminar
Published on July 4, 2022