Contrastive Divergence: Understanding RBM Training

This article provides a clear explanation of the contrastive divergence algorithm for training Restricted Boltzmann Machines (RBMs). By defining the energy function and joint distribution, it derives the weight update rule and explains the role of Gibbs sampling in the positive and negative phases. Ultimately, it shows how the difference between data and model expectations is used to adjust the RBM's weights and biases, minimizing the energy of the training data.