Tutorials
Hands-on technical tutorials. These are full pieces of working code with exercises, not articles. Each tutorial has rendered notebooks here and a source repository with the unfilled exercise stubs (so you can actually run them).
Numerical Linear Algebra for ML - A Hands-On Survey
ARENA-style Jupyter exercises in matrix-free numerical linear algebra, focused on the two objects ML researchers actually care about: the loss Hessian and the empirical NTK. Covers Krylov methods (power iteration, Lanczos with selective reorthogonalization), randomized SVD, stochastic trace and density-of-states estimation, eigenvalue perturbation theory, and a capstone that watches a network's spectrum evolve across training. Assumes no NLA background; assumes comfort with PyTorch and gradient-based ML.
- Krylov methods for the Hessian - HVPs, power iteration, deflation, Lanczos, loss of orthogonality
- Randomized methods and the empirical NTK - stable rank, HMT range finder, matrix-free eNTK matvec
- Trace, density of states, perturbation - Hutchinson, SLQ, Weyl, Davis-Kahan
- Spectroscopy of a network across training - capstone using all of the above