READ: Aggregating Reconstruction Error into Out-of-Distribution Detection

Detecting out-of-distribution (OOD) samples is crucial to the safe deployment of a classifier in real world. However, deep neural networks are known be overconfident for abnormal data. Existing works directly design score function by mining inconsistency from in-distribution (ID) and OOD. In this paper, we further complement with reconstruction error, based on assumption that an autoencoder trained ID data cannot reconstruct OOD as well ID. We propose novel method, READ (Reconstruction Error Aggregated Detector), unify inconsistencies autoencoder. Specifically, error raw pixels transformed latent space classifier. show bridges semantic gap inherits detection performance original. Moreover, adjustment strategy alleviate overconfidence problem according fine-grained characterization Under two scenarios pre-training retraining, respectively present variants our namely READ-MD (Mahalanobis Distance) only pre-trained READ-ED (Euclidean which retrains Our methods do not require access test time fine-tuning hyperparameters. Finally, demonstrate effectiveness proposed through extensive comparisons state-of-the-art algorithms. On CIFAR-10 WideResNet, method reduces average FPR@95TPR up 9.8% compared previous state-of-the-art.