Novelty Detection with Rotated Contrastive Predictive Coding

The current dominant paradigm for novelty detection relies on a learned model’s capability to recover the regularities. To this end, reconstruction-based learning is often used in which the normality of an observation is expressed in how well it can be reconstructed. However, this can be limiting as anomalous data can be reconstructed well if enough common features are shared between normal and anomalous data. In this paper, we pursue an alternative approach wherein the normality is measured by a contrastive learning objective. Specifically, we propose Rotated Contrastive Predictive Coding (Rotated CPC) where the model operates on rotated images and simultaneously learns to predict the future in latent space. Normality score is thus measured as how predictive the representations are and the score’s robustness is further improved by ensembling predictions on multiple rotations of the input signal. We demonstrate the efficacy of this formulation across a variety of benchmark datasets where our method outperforms state-of-the-art methods.