|
I'm trying to implement a Convolutional RBM. I currently implemented a CRBM without probabilistic max pooling following various works of Dr. Honglak Lee. However, I'm not able to reconstruct input from MNIST images. It seems that whatever image I give to the CRBM it gives me almost the same output. I'm almost sure it comes from my implementation, but I would like to know if it is possible to reconstruct input in the same way a RBM does with a CRBM without max-pooling ? |
|
I can finally confirm that it can reconstruct its input :) However, training seems much less stable than a normal RBM. And even for binary units, the learning rate needs to be put low. I probably still have to find better parameters. |
