Machine Learning Research Intern
- We introduce a framework that exploits weight-space symmetries in neural network loss landscapes to construct tractable, structured Hessian approximations from a single gradient, with a user-defined symmetry group controlling the trade-off between accuracy and computational cost.
- The framework unifies existing curvature estimation methods, including Shampoo/Muon-like approaches, and demonstrates strong potential for second-order optimisation and applications such as uncertainty estimation, continual learning, and model compression.
- Check out our ICML paper here.
- Supervisor: Alberto Bernacchia
