A RBM (originally called Harmonium) is a neural model proposed by Smolensky (in Information processing in dynamical systems: Foundations of harmony theory, Smolensky P., Parallel Distributed Processing, Vol 1, The MIT Press) that is made up of a layer of input (observable) neurons and a layer of hidden (latent) neurons. A generic structure is shown in the following diagram:
Structure of Restricted Boltzmann Machine
As the undirected graph is bipartite (there are no connections between neurons belonging to the same layer), the underlying probabilistic structure is MRF. In the original model (even if this is not a restriction), all the neurons are assumed to be Bernoulli-distributed (xi, hi = {0, 1}), with a bias, bi (for the observed units) and cj (for the latent neurons). The resulting energy function is:
A RBM is a probabilistic generative model that can learn a data-generating process, pdata, which is represented by the observed units but exploits the presence of the latent variables...