Weekly summary 2016-09-24

Last week.

Threads

• PMI: WSVD + DL/NMF. PMI for images
  • Used NMF code from here.
• DL generalization
  • Can we tackle NMF using similar algorithm as DL—trying to isolate a column of A by looking at samples which are close to a pair of samples? Before, closeness was inner product; now use another kernel? Ex. TV distance, or KL, or some kind of regularized KL. Ex.
    • In what sense is this important? I would still rely on a distributional assumption on the \(x\)’s. (DL algorithm required this: sparsity and some kind of independence.) But topic modeling already solves this.
    • Why can’t we solve NMF using the topic modeling, saying that the distribution is simply the samples that we get?
• SoS

TODO: run again with 1800-dimensional.

TODO: Review polysemy paper, and understand CKN paper well to rederive equations.