EDGE: Unknown-aware Multi-label Learning by Energy Distribution Gap Expansion

TL;DR

This work tackles multi-label OOD detection and identifies an imbalance in JointEnergy, defined as $\mathcal{E}(\mathbf{z})=\sum_{i=1}^C \log(1+e^{f_{y_i}(\mathbf{x})})$, which narrows the energy gap for tail samples. It introduces EDGE, an unknown-aware multi-label learning framework that combines BCE-based in-distribution learning, an outlier-exposure term $\mathcal{L}_{\text{conf}}$, and an energy-gap term $\mathcal{L}_{\text{gap}}$, with a loss $\mathcal{L}_{\text{edge}}=\mathcal{L}_{\text{id}}+\alpha\mathcal{L}_{\text{conf}}+\beta\mathcal{L}_{\text{gap}}$, to enlarge the separation between tail-ID and OE. A feature-based OE selection via the dilation distance $d_{\text{dilation}}=\lVert \Sigma_{\text{in}}^k-\Sigma_{\text{out}}^k\rVert_F$ guides informative auxiliary data. Comprehensive experiments on PASCAL-VOC, MS-COCO, and NUS-WIDE with various OOD datasets demonstrate state-of-the-art OOD detection performance while preserving ID accuracy, validating the effectiveness of energy-space reshaping for multi-label OOD detection.

Abstract

Multi-label Out-Of-Distribution (OOD) detection aims to discriminate the OOD samples from the multi-label In-Distribution (ID) ones. Compared with its multiclass counterpart, it is crucial to model the joint information among classes. To this end, JointEnergy, which is a representative multi-label OOD inference criterion, summarizes the logits of all the classes. However, we find that JointEnergy can produce an imbalance problem in OOD detection, especially when the model lacks enough discrimination ability. Specifically, we find that the samples only related to minority classes tend to be classified as OOD samples due to the ambiguous energy decision boundary. Besides, imbalanced multi-label learning methods, originally designed for ID ones, would not be suitable for OOD detection scenarios, even producing a serious negative transfer effect. In this paper, we resort to auxiliary outlier exposure (OE) and propose an unknown-aware multi-label learning framework to reshape the uncertainty energy space layout. In this framework, the energy score is separately optimized for tail ID samples and unknown samples, and the energy distribution gap between them is expanded, such that the tail ID samples can have a significantly larger energy score than the OOD ones. What's more, a simple yet effective measure is designed to select more informative OE datasets. Finally, comprehensive experimental results on multiple multi-label and OOD datasets reveal the effectiveness of the proposed method.

Figures (5)

• Figure 1: Distribution and performance results with Binary Cross-Entropy loss (BCE) training loss and our framework on Places50 dataset. (a) shows the small gap between known and unknown domains under basic multi-label models, and (b) indicates the improvement of our method. The last two figures show the (c) FPR95 and (d) AUROC imbalanced and balanced results on different in-distribution subsets, where the proportion of in-distribution samples that only relate to minority classes increases with the x-axis value increases.
• Figure 2: Overview of EDGE learning framework. This process could be divided into two parts, i.e., auxiliary OE data selection, and EDGE loss optimization. Firstly, the candidate OE datasets are fed into the OE Selection Module (OESM, shown in Fig. \ref{['fig:osm']}). Then, the deep model is provided with both in-distribution data and an informative auxiliary dataset to conduct unknown-aware multi-label learning with the two-part EDGE loss.
• Figure 3: Illustration of Outlier Exposure Selection Module.
• Figure 4: UMAP visualization of feature embedding results.
• Figure 5: Performance change curve on PASCAL-VOC. The closer to the right, the larger the proportion of tail samples.