Adequate structuring of the latent space for easy classification and out-of-distribution detection

Out-of-distribution (OoD) detection is the cornerstone of reliability in machine learning (ML) applications. Since OoD samples follow a different statistic than those on which the model is trained, the corresponding model decision is likely to be unreliable, and OoD samples must be identified as such. Moreover, OoD samples can follow any statistic, which calls for an unsupervised method (independent of the OoD statistics). It is already well known that variational auto-encoder (VAE) based classification can be improved by structuring the latent space in terms of the class centroids. In this paper, we extend this approach by adding an appropriate structure to the latent space for OoD detection. The corresponding performance is precisely analysed, demonstrating the benefits of the approach.