Perceptual Regularization: Visualizing and Learning Generalizable Representations

A deployable machine learning model relies on a good representation. Two desirable criteria of a good representation are to be understandable, and to generalize to new tasks. We propose a technique termed perceptual regularization that enables both visualization of the latent representation and control over the generality of the learned representation. In particular our method provides a direct visualization of the effect that adversarial attacks have on the internal representation of a deep network. By visualizing the learned representation, we are also able to understand the attention of a model, obtaining visual evidence that supervised networks learn task-specific representations. We show models trained with perceptual regularization learn transferrable features, achieving significantly higher accuracy in unseen tasks compared to standard supervised learning and multi-task methods.