Adversarial Adaptive Interpolation in Autoencoders for Dually Regularizing Representation Learning

Linear interpolation in latent space may induce mismatch between the constructed data and the distribution a model was trained on. In this paper, we propose a dually regularized AutoEncoder-based representation learning model with Adversarial Adaptive Interpolation, which is referred to as AdvAI-AE. To constrain the interpolation path on the underlying manifold, an additional interpolation correction module is trained to offset the deviation between the linearly interpolated data points and the statistics of real ones in latent space. Further, we apply prior matching to control the characteristics of the representation. Toward this end, the maximum mean discrepancy-based and adversarial regularizers are incorporated into the model. The synthesized data from random variables are in turn leveraged to regularize the interpolation process. The proposed improvement strategies leads to significant performance gains in downstream classification, clustering and synthesis tasks on multiple benchmark datasets.