how to design feature extraction layer for DBN for face recognition

Question

I am trying to use deep belief networks for face recognition. But I am a beginner in this area, I have read the research papers and documentations available on the Internet and I understood the basic concept for binary images. But still when I sit down to code I find great difficulty because nothing is explained from a programmers perspective all you find is energy functions and all that stuff. Can some body help me design(code) hidden layer for a gray scale Face image ? (To be more specific what should my hidden layer be should it be an array of different filters or something else ....)

Answer 1

I have experience with Neural Networks and Self-Organizing Maps dating back to the late-1980s, but I too find Energy-based Restricted Boltzmann Machines somewhat daunting to just sit down and implement. I found the following websites with either Matlab code (Octave?) and C. They're from the Netflix competition (winner from University of Toronto), but it's a good winning example and should provide some insight.

http://imonad.com/rbm/restricted-boltzmann-machine/

https://code.google.com/p/nprizeadditions/source/browse/trunk/rbm.c

I might also suggest taking Andrew Ng's Coursera on Machine Learning from Stanford (it's free and a new session starts on Jan 19, 2015) I've viewed a number of the lectures and they are very good. Hope this helps.