II. DEEP BELIEF NETWORKSHinton et al. proposed a DBN as an efficient unsupervisedlearning algorithm to overcome the complexity of trainingdeep generative model [12]. DBN is a hierarchical structureof multiple layers of restricted Boltzmann machines (RBM).The process of training DBNs involves the individual trainingof each RBM one after another and then stack them on topof each other. The outputs of the hidden nodes at layer l − 1is used as input data for training the next RBM at layer l. Ineach RBM, there is neither visible-visible nor hidden-hiddenconnection but there are direct connections between units ofvisible and hidden layer. RBM consists of a weight wij , wherei represents a visible node and j represents a hidden node, avisible bias ai of node i, and a hidden bias bi of node j.Gaussian or Bernoulli stochastic visible units are usually usedin RBMs while the hidden units are usually Bernoulli [12].For a Bernoulli(visible)-Bernoulli(hidden) RBM, the energyfunction of a joint configuration (v, h) of the visible andhidden units is described by: