UCLA

In this work, we use a 2D to study many aspects of Deep FRAME model. We first do visualization on the training process, showing how the fitted probability distribution evolves during training, how the model captures different modes and how these results are influenced by the choice of prior distributions and activation functions. We then study the activation pattern of the learned network and the corresponding partition of the input space. Based on the input space partition and statistics matching, we then compare the multi-layer model trained with SGD with the original one-layer model. Finally, we analyze the case in which we train the model with finite-MCMC sampling, showing the difference between tting the energy function and synthesizing samples.