Interpretable Recurrent Neural Networks for reconstructing nonlinear dynamical systems from time series observations

About

Mathematical models of natural processes in physics, biology, neuroscience, and beyond, are commonly formulated in terms of differential or time-recursive equations, i.e. dynamical systems. Traditionally, such models are constructed in a 'top-down' fashion, i.e., conceived by theoreticians and then refined in iterations with experimental results. Modern machine learning tools may help to augment this process by 'bottom-up', strongly data-driven strategies. The question here is: Given a set of time series observations from some physical or biological system, can we infer from these observations alone the underlying dynamical system, or governing equations, that gave rise to them? My talk will review recent methodological advances toward this goal, based on deep recurrent neural networks (RNNs) as universal approximators of dynamical systems. Some methodological issues, examples on common benchmark ('ground truth') dynamical systems, applications in neuroscience, and analysis of inferred RNNs, will be discussed.

Colloquium 15 July 2021: https://cern.zoom.us/j/64971800448?pwd=Z3FlNUE2bHdoVHhra0lTdTZ4RTJxQT09.

0:00:00 Slide 1

0:07:38 Slide 2

0:15:16 Slide 3

0:22:53 Slide 4

0:30:31 Slide 5

0:38:09 Slide 6

0:45:46 Slide 7

0:53:24 Slide 8

1:01:01 Slide 9

1:08:39 Slide 10