Understanding Minimum Probability Flow for RBMs Under Various Kinds of Dynamics

Energy-based models are popular in machine learning due to the elegance of their formulation and their relationship to statistical physics. Among these, the Restricted Boltzmann Machine (RBM) has been the prototype for some recent advancements in the unsupervised training of deep neural networks. However, the contrastive divergence training algorithm, so often used for such models, has a number of drawbacks and ineligancies both in theory and in practice. Here, we investigate the performance of Minimum Probability Flow learning for training RBMs. This approach reconceptualizes the nature of the dynamics defined over a model, rather than thinking about Gibbs sampling, and derives a simple, tractable, and elegant objective function using a Taylor expansion, allowing one to learn the parameters of any distribution over visible states. In the paper, we expound the Minimum Probability Flow learning algorithm under various dynamics. We empirically analyze its performance on these dynamics and demonstrate that MPF algorithms outperform CD on various RBM configurations.