On Tue, Dec 5, 2017 at 7:36 PM Devon McCormick wrote:
Thanks for the image-processing suggestions. I have lots of pictures - "million" is not an exaggeration - and I'd like to do some kind of automated recognition and aesthetic ranking of them.
Big data problem for sure. Dig around in locality hashing. Or if you have bunches of GPUs, you could try to cook up a J native DL framework using Pascal's Arrayfire thingee. You might get lucky and be able to pull out super simple features like a histogram of the colors, the gradient, maybe some haar filtered numbers showing gross patterns and so on, and use a simple classifier to tell them apart. I never wrote a 2-d wavelet for J, but it's a simple extension to the existing 1-d code.
"AI" and machine learning is a lot like being a chef. You need lots of ingredients to make it work.
FWIIW Don; it is quadratic programming; you're solving giant systems of coupled quadratic equations, with constraints. You can jury rig Simplex to do this (I think there is a basic Simplex solver in addons/math somewhere). I would admire anyone who can do this; I wrote one once (back in dark ages before open source Simplex solvers), but I totally wouldn't advise it unless you are really curious about quadratic programming!
-SL ---------------------------------------------------------------------- For information about J forums see http://www.jsoftware.com/forums.htm
