Isotropic Maximization Loss and Entropic Score: Accurate, Fast, Scalable, Turnkey, and Native Neural Networks Out-of-Distribution Detection

Current out-of-distribution detection (ODD) approaches require cumbersome procedures that add undesired side effects to the solution. In this paper, we argue that the low ODD performance of neural networks is mainly due to SoftMax loss anisotropy. Consequently, we propose an isotropic loss (IsoMax) and a score (Entropic Score) to significantly improve their ODD performance while keeping the overall solution accurate, fast, scalable, turnkey, and native. Our experiments indeed confirmed that neural networks ODD performance may be extremely improved simply by replacing the SoftMax loss without relying on techniques such as adversarial training or validation, special-purpose data augmentation, outlier exposure, ensembles methods, generative approaches, model architectural changes, metric learning, or additional classifiers or regressions. The results also showed that our straightforward proposal is competitive against state-of-the-art approaches besides avoiding their undesired requirements and weaknesses.