NashAE: Disentangling Representations Through Adversarial Covariance Minimization

We present a self-supervised method to disentangle factors of variation in high-dimensional data that does not rely on prior knowledge the underlying profile (e.g., no assumptions number or distribution individual latent variables be extracted). In this which we call NashAE, feature disentanglement is accomplished low-dimensional space standard autoencoder (AE) by promoting discrepancy between each encoding element and information recovered from all other elements. Disentanglement promoted efficiently framing as minmax game AE an ensemble regression networks provide estimate conditioned observation quantitatively compare our approach with leading methods using existing metrics. Furthermore, show NashAE has increased reliability capacity capture salient characteristics learned representation.