Neural nets learn dictionaries

Posted: 2016-10-10 , Modified: 2016-10-10

Tags: dictionary learning, neural nets

writeup

Result

General idea: when each row of \(B\) is sufficiently close to the row of \(A^T\), gradient descent converges until \(B\) is very close to \(A^T\).

PROBLEM: I’ve done a lot of the calculations for neural nets learning dictionaries, and am getting stuck on the following: it appears that the gradient of the entire matrix is correlated with the right direction, but individual rows may not be (so a row may get far away until it no longer decodes correctly).

Proof

Further directions