Top-down processing of noisy and adversarial stimuli in deep generative models

This thesis studies the effect of adding a term usually neglected during the training phase of energy-based models. It is called ``KL term'' because of its dependence on Kullback-Leibler divergence, and it does not have a significant impact on the training in terms of running time and computational cost. I will initially present an analysis of its impact on training stability and some considerations relative to the general structure of the learning model. I will then study the denoising capabilities of the model by implementing top-down processes applied to different types of noisy input. Thirdly, to understand the quality of internal representations emerging in the hidden layers, I will apply a read-out classifier to the deepest hidden layer, calculating psychometric curves produced with different noise values. The final analysis will explore the model capability to resist to adversarial attacks using forward-backward iterations, also considering the spontaneous generative activity of the network by stimulating the read-out neurons relative to a specific class. Interestingly, results of this latter investigation are very encouraging for the MNIST dataset, suggesting that this type of energy-based models has potential to improve current defenses on adversarial attacks. However, for Cifar10 dataset it seems that more powerful computing hardware is needed to train models of larger sizes