How to test a Restricted Boltzmann Machine implementation ?

Question

I developed a simple binary Restricted Boltzmann Machine implementation and now I would like to test it. (Ultimately I'm gonna use it for a DBN, but I would like to test independently).

I saw that several people and papers are talking about testing it MNIST dataset, but I didn't found details on how to do that.

Do I have to add a new classification layer connected to the hidden units and then use back propagation to train it ? Isn't there another way ?

Some people are also plotting the weights (again in MNIST), but I have problems on how you can plot a weight and what does that represent...

Thanks

Answer 1

The "Tracking Progress" section in the RBM tutorial at deeplearning.net (http://deeplearning.net/tutorial/rbm.html) gives very good guidance:

1. Check that samples from the RBM look like the training data
2. (For image data) Check that latent variable values maxima look sort of like smooth gabor filter banks
3. Track the pseudolikelihood