Training Reject-Classifiers for Out-of-distribution Detection via Explicit Boundary Sample Generation

Discriminatively trained neural classifiers can be trusted only when the input data comes from the training distribution (in-distribution). Therefore, detecting out-of-distribution (OOD) samples is very important to avoid classification errors. In the context of OOD detection for image classification, one of the recent approaches proposes training a classifier called \“confident-classifier\” by minimizing the standard cross-entropy loss on in-distribution samples and minimizing the KL divergence between the predictive distribution of OOD samples in the low-density \“boundary\” of in-distribution and the uniform distribution (maximizing the entropy of the outputs). Thus, the samples could be detected as OOD if they have low confidence or high entropy. In this work, we analyze this setting both theoretically and experimentally. We also propose a novel algorithm to generate the \“boundary\” OOD samples to train a classifier with an explicit \“reject\” class for OOD samples. We show that this approach is effective in reducing high-confident miss-predictions on OOD samples while maintaining the test-error and high-confidence on the in-distribution samples compared to standard training. We compare our approach against several recent classifier-based OOD detectors including the confident-classifiers on MNIST and FashionMNIST datasets. Overall the proposed approach consistently performs better than others across most of the experiments.