search

amaas / stanford_dl_ex

Programming exercises for the Stanford Unsupervised Feature Learning and Deep Learning Tutorial
http://ufldl.stanford.edu/tutorial
MIT License
2.59k stars 1.6k forks source link

RICA exercise: what should "sane results" look like? #13

Open ypeels opened 9 years ago

ypeels commented 9 years ago

Does anybody have an example of the "sane results" mentioned in runSoftICA.m? Looks like image captions aren't supported, so I'll limit things to one image per post for legibility... For completeness, here is my source code...

Here's what I am getting for the default parameter values (10000 patches). normalized

ypeels commented 9 years ago

Default parameter values (10000 patches) but without "Step 3) Normalize each patch", instead just setting x = patches. The reason for doing this is given 2 posts below. unnormalized

ypeels commented 9 years ago

200,000 patches + 32 filters (as in stlExercise.m) + normalization. Things look "over-regularized"... stock-normalized

ypeels commented 9 years ago

200,000 patches + 32 filters (as in stlExercise.m) + no normalization.

This was the only way I've found so far to obtain filters resembling the tutorial's (Step 2). I can get them to look somewhat more (qualitatively) like the tutorial's by tweaking lambda, but the result shown below is for default parameter values. stock-unnormalized

ypeels commented 9 years ago

An alternative? 200,000 patches + 32 filters + normalization + epsilon = 0.1 in zca2.m: untitled