Intra-class Mixup for Out-of-Distribution Detection

Deepak Ravikumar, Sangamesh Kodge, Isha Garg, Kaushik Roy

Abstract: Deep neural networks have found widespread adoption in solving image recognition and natural language processing tasks. However, they make confident mispredictions when presented with data that does not belong to the training distribution, i.e. out-of-distribution (OoD) samples. Inter-class mixup has been shown to improve model calibration aiding OoD detection. However, we show that both empirical risk minimization and inter-class mixup create large angular spread in latent representation. This reduces the separability of in-distribution data from OoD data. In this paper we propose intra-class mixup supplemented with angular margin to improve OoD detection. Angular margin is the angle between the decision boundary normal and sample representation. We show that intra-class mixup forces the network to learn representations with low angular spread in the latent space. This improves the separability of OoD from in-distribution examples. Our approach when applied to various existing OoD detection techniques shows an improvement of 4.68% and 6.38% in AUROC performance over empirical risk minimization and inter-class mixup, respectively. Further, our approach aided with angular margin improves AUROC performance by 7.36% and 9.10% over empirical risk minimization and inter-class mixup, respectively.