Noether Networks - Meta-Learning Useful Conserved Quantities

Explore the concept of Noether Networks in this in-depth video featuring an interview with first author Ferran Alet. Delve into the innovative approach of using Noether's theorem to connect symmetries with conserved quantities, enabling the dynamic and approximate enforcement of symmetry properties in deep neural networks. Learn about the potential of automatically discovering useful symmetries to improve machine learning performance, with a focus on sequential prediction problems. Gain insights into the architecture of Noether Networks, the optimization of Noether Loss, and the framework for discovering inductive biases. Examine experimental results and engage in a thoughtful discussion on the implications and applications of this groundbreaking research in the field of deep learning and symmetry exploitation.

- Intro & Overview
- Interview Start
- Symmetry priors vs conserved quantities
- Example: Pendulum
- Noether Network Model Overview
- Optimizing the Noether Loss
- Is the computation graph stable?
- Increasing the inference time computation
- Why dynamically modify the model?
- Experimental Results & Discussion