COLLOQUIUM 2025

Abstract

How much can we learn about a dynamical system when models are unavailable and data is limited? In the first half of this talk, I will pose basin prediction in multistable systems as a challenging task to probe out-of-domain generalization in neural networks. Surprisingly, reservoir computers (a class of simplified recurrent neural networks) without relevant physics baked in can often extrapolate and reconstruct basins not seen during training. In the second half of the talk, I will discuss time-series foundation models’ ability to forecast a new dynamical system based on a short context trajectory. We find that foundation models excel at zero-shot forecasting and often utilize context parroting as a key mechanism to forecast chaotic systems. Exploring zero-shot forecasting strategies beyond parroting can provide insights into how artificial and natural intelligence learn from limited data.