*** DBN_125.mat ***

Deep Belief Network trained on the MNIST data set in a completely unsupervised way, using greedy layer-wise contrastive divergence.
Network architecture: 500-500-2000 (three layers).


*** DBN_125_sparse.mat ***

Deep Belief Network trained on the MNIST data set in a completely unsupervised way, using greedy layer-wise contrastive divergence.
Network architecture: 500-500-2000 (three layers).
A sparsity constraint has been introduced in the third hidden layer in order to encourage the emergence of more structured receptive fields and reduce the number of dead units.
