Comments | markgritter: Principal Component Analysis

markgritter

Principal Component Analysis

Oct 25, 2016 20:42

A similar technique to NNMF from my last entry is Principal Component Analysis. It still factors the input matrix, but now negative values are allowed. A big challenge for using PCA is what to do with those negative numbers, which have no standard interpretation. (This is one of the advantages of NNMF, that it can be applied in scenarios in ( Read more... )

geek, programming

Comments 3

markgritter October 26 2016, 02:18:48 UTC

Trying it on a larger set produces results that are... not boring, at least?

( ... )

markgritter October 26 2016, 02:23:31 UTC

It might be better to do PCA on colors individually, or red/green/blue/alpha color channels individually.

markgritter October 26 2016, 03:07:27 UTC

Sparse PCA seemed like an appealing thing to try because it promised components that were individually composed of few pixels. However, the pixels seem to be random and even at n=100 do a poor job reconstructing the originals:

( ... )