When training a RBM how should the data be normalised?

A feed forward neural network trains best if the data is normalised so that each input has a mean of 0 and a standard deviation of 1. Is this true for a Restricted Boltzmann machine as well? (My experiments are saying, no it isn't.)

• Based on my experience and some papers if you are working on binary RBM you should normalize your data between [0,1]. could you explain why your experiment says no? – Feras Jul 9 '16 at 9:33

That depends on the distributions you assume the visible and hidden variables to have. If the visibles are binary, you should scale them to the [0, 1] interval. If they are Gaussian, you should scale them to $\mathcal{N}(0, 1)$.