Out-of-distribution Detection with Boundary Aware Learning

AbstractThere is an increasing need to determine whether inputs are out-of-distribution (OOD) for safely deploying machine learning models in the open world scenario. Typical neural classifiers based on closed assumption, where training data and test drawn i.i.d. from same distribution, as a result, give over-confident predictions even faced with OOD inputs. For tackling this problem, previous studies either use real outliers or generate synthetic under strong assumptions, which costly intractable generalize. In paper, we propose boundary aware (BAL), novel framework that can learn distribution of features adaptively. The key idea BAL trivial hard progressively generator, meanwhile, discriminator trained distinguishing these in-distribution (ID) features. Benefiting adversarial scheme, well separate ID features, allowing more robust detection. proposed achieves state-of-the-art performance classification benchmarks, reducing up 13.9% FPR95 compared methods.KeywordsOOD detectionBoundary learningGAN