A Cyclically-Trained Adversarial Network for Invariant Representation Learning

Recent studies show that deep neural networks are vulnerable to adversarial examples which can be generated via certain types of transformations. Being robust to a desired family of adversarial attacks is then equivalent to being invariant to a family of transformations. Learning invariant representations then naturally emerges as an important goal to achieve which we explore in this paper within specific application contexts. Specifically, we propose a cyclically-trained adversarial network to learn a mapping from image space to latent representation space and back such that the latent representation is invariant to a specified factor of variation (e.g., identity). The learned mapping assures that the synthesized image is not only realistic, but has the same values for unspecified factors (e.g., pose and illumination) as the original image and a desired value of the specified factor. Unlike disentangled representation learning, which requires two latent spaces, one for specified and another for unspecified factors, invariant representation learning needs only one such space. We encourage invariance to a specified factor by applying adversarial training using a variational autoencoder in the image space as opposed to the latent space. We strengthen this invariance by introducing a cyclic training process (forward and backward cycle). We also propose a new method to evaluate conditional generative networks. It compares how well different factors of variation can be predicted from the synthesized, as opposed to real, images. In quantitative terms, our approach attains state-of-the-art performance in experiments spanning three datasets with factors such as identity, pose, illumination or style. Our method produces sharp, high-quality synthetic images with little visible arte-facts compared to previous approaches.