Module I: Infinite-width networks. (2 weeks)

• Kernels, Symmetries and Representations, Spectral Bias in supervised tasks and diffusion models, Effective Ridge.

Module II: Feature Learning. (2 weeks)

• Kernel adaptation, grokking, phase transitions, kernel spikes and sample complexity changes.

Module III: Scaling behavior. (3 weeks)

• Data scaling laws, Neural scaling laws, feature scaling laws, hyperparameter transfer, fine-tuning scaling laws and algorithmic capture.

Module IV: TBD. (1 week)

• [Renormalization-group, spin-glass theory for RLVR].

-----

This mini-course will take place in-person at Room 210, Fields Institute and online via Zoom.

Optional homework assignments will be given out weekly, but there will be no formal evaluation.