 |
Center for Nonlinear Studies |
Reservoir computing (RC) has been shown to be a simple and effective framework to learn dynamics from data. Most work on RC has focused on learning systems with a single chaotic attractor. In this talk, I will show that learning multistable systems (e.g., reconstructing basins of attraction) poses unique challenges for both traditional RC and next-generation reservoir computing (NGRC) frameworks. For NGRC, I will also discuss another curious property: When given more (noise-free) training data, sometimes the prediction of NGRC becomes much worse.
Join by phone
+1-415-655-0002 US Toll
Access code: 2631 343 9655
Host PIN: 3793
Host: Melvyn Tyloo (T-4)