Skip to yearly menu bar Skip to main content


Poster
in
Workshop: Machine Learning and the Physical Sciences

Noether Networks: Meta-Learning Useful Conserved Quantities

Ferran Alet · Dylan Doblar · Allan Zhou · Josh Tenenbaum · Kenji Kawaguchi · Chelsea Finn


Abstract:

Progress in machine learning (ML) relies on an appropriate encoding of inductive biases. Useful biases often exploit symmetries in the prediction problem, such as convolutional networks relying on translation equivariance. Automatically discovering these useful symmetries holds the potential to greatly improve the performance of ML systems, but still remains a challenge. In this work, we focus on sequential prediction problems and take inspiration from Noether's theorem to reduce the problem of finding inductive biases to meta-learning useful conserved quantities. We propose Noether Networks: a new type of architecture where a meta-learned conservation loss is optimized inside the prediction function. We show, theoretically and experimentally, that Noether Networks improve prediction quality, providing a framework for discovering inductive biases in sequential problems.

Chat is not available.