Improving Energy-Based Out-of-Distribution Detection by Sparsity Regularization

Abstract

Out-of-distribution (OOD) detection is critical for safely deploying machine learning models in the open world. Recently, an energy-score based OOD detector was proposed for any pre-trained classification models. The energy score, which is less susceptible to overconfidence, proves to be a better substitute for the conventional approaches leveraging the softmax confidence score. However, current energy-score based methods rely heavily on large-scale auxiliary datasets and introduce several dataset-dependent hyperparameters. In this paper, we propose a simple yet effective sparsity-regularized learning objective for deep neural networks so that the energy-based detector works better. Our learning objective is parameter-free and its key idea is to enlarge the differences between network outputs of in-distribution data and OOD data by regularizing the networks to generate high sparsity representations for in-distribution data. We also contribute to a tiny auxiliary outlier dataset to replace the previous one, which reduces the volume size significantly (230G vs. 40M). Besides, a new energy-score based OOD detector named Sparsity-Regularized Outlier Exposure (SROE) is proposed to incorporate the proposed sparsity-regularized loss function into the traditional Outlier Exposure method. Experimental results show that the proposed sparsity-regularized loss strategy is effective, and the SROE OOD detector outperforms the other SOTA methods with a large margin. The source code and dataset are available at https://github.com/kuan-li/SparsityRegularization.