Using a restricted Boltzmann machine to reconstruct Bangla MNIST images

A restricted Boltzmann machine (RBM) is an unsupervised model. As an undirected graphical model with two layers (observed and hidden), it is useful to learn a different representation of input data along with the hidden layer. This was the first structural building block of deep learning, particularly when the computational resources to learn about a deep neural net with backpropagation were not available (a stacked RBM was used instead). It restricts the connectivity of the network (only allowing a bipartite graph in between the hidden and observed set of nodes) to make inference easy. It is an energy-based model; the joint distribution is modeled using the energy function. To infer the most probable observation, we need to choose the one with the least energy. This model is generally trained on binary images...