Today

  • Eigen-decomposition and SVD (cont’d).

  • Optimization introduction.

Announcement

  • HW2 graded. Common issues:
    • Late submission. Remember to git tag.
    • NNMF: local minimum due to non-convexity; V * W * W' vs V * (W * W'); use in-place functions mul! to reduce allocation of intermediate arrays; basis images show parts of face; GPU coding.
    • Multivariate normal density. Using Woodbury structure should see ~100 fold speedup compared to standard way to evaluate normal density.

    You can re-work on homework problems and re-submit for bonus points.

  • HW3 due Wed 5/15 @ 11:59PM. PageRank problem: it took me <6 seconds to obtain solution using GMRES and <16 seconds using Arnoldi. The key is to use the LinearMaps.jl to define fast matrix-vector multiplication utilizing the Woodbury structure (sparse + rank 1).

  • HW4 posted (SVD). Due Fri 5/24 @ 11:59PM.

  • Next Tue (5/14)’s class ends at 1:45p. Xiao-Li Meng’ talk 2p-3:15p.