Noether Networks: Meta-Learning Useful Conserved Quantities
This work addresses the problem of enhancing machine learning performance through automated inductive bias discovery for researchers and practitioners in sequential prediction domains, representing a novel approach rather than an incremental improvement.
The paper tackles the challenge of automatically discovering useful symmetries in sequential prediction problems by introducing Noether Networks, a meta-learning architecture that optimizes a meta-learned conservation loss, and demonstrates improved prediction quality both theoretically and experimentally.
Progress in machine learning (ML) stems from a combination of data availability, computational resources, and 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 general framework for discovering inductive biases in sequential problems.