|
I'm having trouble reproducing Figure 3 in H. Lee's ICML 09 paper on convolutional RBMs. I can train the first layer of the DBN and get similar results to the paper, but I can't seem to train the second layer properly. Whatever I do, the weights don't look like "parts of faces". I have modified H. Lee's Matlab code using information I found in this forum post, but I'm still not convinced it's correct. Equally, the code could be correct, and the problem lies with the input parameters. Does anyone have a working version of the convolutional deep belied network code, or can offer any insight into what I might be doing wrong? Thanks in advance - I've really been struggling with this one! |
