Multi-Label Out-of-Distribution Detection with Spectral Normalized Joint Energy

TL;DR

This work tackles multi-label out-of-distribution (OOD) detection by introducing Spectral Normalized Joint Energy (SNoJoE), which aggregates label-specific information through an energy-based framework and enforces stability via spectral normalization. By defining a label-wise joint energy $E_{joint}(x)=\sum_{i=1}^K -E_{y_i}(x)$ with $E_{y_i}(x)=-\ln(1+e^{f_{y_i}(x)})$ and applying bi-Lipschitz constraints to the initial network layers, the method achieves robust OOD discrimination. Empirical results on PASCAL-VOC as ID with ImageNet-22K and Texture as OOD demonstrate state-of-the-art performance, including 11% and 54% relative reductions in FPR95 over prior approaches and high AUROC/AUPR scores, validating both the approach and the importance of spectral normalization in multi-label OOD tasks. The paper also provides extensive ablations showing that carefully positioned spectral normalization improves generalizable feature extraction, and it releases code and datasets to promote reproducible research. Overall, SNoJoE advances multi-label OOD detection by combining energy-based uncertainty with spectral regularization, offering practical gains for reliable in/deployment in open-world settings.

Abstract

In today's interconnected world, achieving reliable out-of-distribution (OOD) detection poses a significant challenge for machine learning models. While numerous studies have introduced improved approaches for multi-class OOD detection tasks, the investigation into multi-label OOD detection tasks has been notably limited. We introduce Spectral Normalized Joint Energy (SNoJoE), a method that consolidates label-specific information across multiple labels through the theoretically justified concept of an energy-based function. Throughout the training process, we employ spectral normalization to manage the model's feature space, thereby enhancing model efficacy and generalization, in addition to bolstering robustness. Our findings indicate that the application of spectral normalization to joint energy scores notably amplifies the model's capability for OOD detection. We perform OOD detection experiments utilizing PASCAL-VOC as the in-distribution dataset and ImageNet-22K or Texture as the out-of-distribution datasets. Our experimental results reveal that, in comparison to prior top performances, SNoJoE achieves 11% and 54% relative reductions in FPR95 on the respective OOD datasets, thereby defining the new state of the art in this field of study.