Icml2022 | partial and asymmetric comparative learning of out of distribution detection in long tail recognition

Outside the existing distribution (OOD) The detection method is usually based on the training set with balanced class distribution . However , in application , Training sets usually have long tailed distributions . In this work , We first proved the existing OOD When the training set is a long tailed distribution, the performance of the detection method usually decreases significantly . Through analysis , We think this is because it is difficult for the model to compare a few tailed samples in the distribution with the real OOD Samples are distinguished , It makes the tail class easier to be incorrectly detected as OOD. To solve this problem , Partial and asymmetric supervised comparative learning (Partial and Asymmetric Supervised contrast Learning, PASCL), The model is explicitly encouraged to distinguish between tail intraclass distribution samples and OOD sample . In order to further improve the classification accuracy within the distribution , We propose auxiliary branch tuning (Auxiliary Branch Finetuning), It USES BN And two independent branches of the classification layer for anomaly detection and intra distribution classification . Intuition , Abnormal data and OOD Abnormal data has different underlying distributions . Our approach is CIFAR10-LT、CIFAR100-LT and ImageNet-LT False positive rate of abnormal detection on (FPR) Respectively 1.29%、1.45%、0.69%, The classification accuracy within the distribution is 3.24%、4.06%、7.89%. Code and pre trained models can be found in https://github.com/amazon-research/long-tailed-ood-detection find .