MIT Logo

Earth Signals and Systems Group

Informatics and Intelligent Systems for Earth, Atmospheric, and Planetary Science

The Learnability of Feedforward Neural Networks in Dissipative Chaotic Dynamical Systems

Thursday, November 14, 2019

Ziwei Li and Sai Ravela discuss how

  • The predictability of neural networks is near identical to the chaotic dynamics they learn from
  • This they argue is because neural Networks follow a path to chaos through alternate stretching and compression
  • Very few training points are needed to learn the Lorenz attractor

See you at AGU 2019!