It's common to make visualizations of the features that hidden units learn on image datasets as we can gain good intuition about what's happening during learning. For gray scale datasets, it's easy to use a gray scale color-map, but for color datasets, the proper procedure seems to be a lot less clear.

Anyone have knowledge of the common/standard ways of taking your raw weights (negative and positive values for each color channel over semi-arbitrary ranges) and visualizing them as color images?

asked Oct 20 '13 at 18:01

Newmu's gravatar image

Newmu
29641014

edited Oct 20 '13 at 18:02

Alex Krizhevsky's master thesis (http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.222.9220&rep=rep1&type=pdf), has a paragraph on the approach used there:

"In creating these visualizations, we use intensity to indicate the strength of the weight. So a green region denotes a set of weights that have strong positive connections to the green channel of the pixels. But a purple region indicates strong negative connections to the green channel of the pixels, since this is the colour produced by setting the red and blue channels to high values and the green channel to a low value. In this sense, purple is negative green, yellow is negative blue, and turquoise is negative red."

(Oct 21 '13 at 22:08) Newmu

Anyone else have input? His approach seems to look okay, but I'm wondering if there are others.

(Oct 26 '13 at 16:23) Newmu
Be the first one to answer this question!
toggle preview

powered by OSQA

User submitted content is under Creative Commons: Attribution - Share Alike; Other things copyright (C) 2010, MetaOptimize LLC.