Adaptive Negative Evidential Deep Learning for Open-set Semi-supervised Learning

TL;DR

Open-set SSL faces the challenge of unlabeled data containing novel categories. The paper proposes Adaptive Negative Evidential Deep Learning (ANEDL), which combines an evidential outlier detector with a traditional Softmax classifier and introduces adaptive negative optimization to suppress outlier evidence while preserving inlier learning. It also defines dedicated uncertainty metrics for self-training ($M_{Self-training}$) and inference ($M_{Inference}$), leveraging Fisher information via Dirichlet evidence to guide learning and outlier detection. Empirically, ANEDL achieves state-of-the-art AUROC and lower error rates on CIFAR-10/100, ImageNet-30, and Mini-ImageNet, particularly when the inlier class set is large. The approach enhances scalability and robustness of open-set SSL, enabling better utilization of unlabeled data in real-world settings.

Abstract

Semi-supervised learning (SSL) methods assume that labeled data, unlabeled data and test data are from the same distribution. Open-set semi-supervised learning (Open-set SSL) considers a more practical scenario, where unlabeled data and test data contain new categories (outliers) not observed in labeled data (inliers). Most previous works focused on outlier detection via binary classifiers, which suffer from insufficient scalability and inability to distinguish different types of uncertainty. In this paper, we propose a novel framework, Adaptive Negative Evidential Deep Learning (ANEDL) to tackle these limitations. Concretely, we first introduce evidential deep learning (EDL) as an outlier detector to quantify different types of uncertainty, and design different uncertainty metrics for self-training and inference. Furthermore, we propose a novel adaptive negative optimization strategy, making EDL more tailored to the unlabeled dataset containing both inliers and outliers. As demonstrated empirically, our proposed method outperforms existing state-of-the-art methods across four datasets.

Figures (4)

• Figure 1: Overview of our proposed Adaptive Negative Evidential Deep Learning (ANEDL). (a) The framework of ANEDL consists of a shared feature extractor, a Softmax head and an EDL head. EDL head is used to detect outliers, while Softmax head is used to classify. To effectively leverage the information of inlier and outlier from unlabeled samples, we introduce Negative Optimization to explicitly regularize our EDL detector to output low evidence values for uncertain classes, while proposing adaptive loss weights to encourage the detector to pay more attention to these uncertain classes. (b) Three cases of unlabeled samples. Our model can quantify different types of uncertainty, including epistemic uncertainty due to lack of knowledge, and aleatoric uncertainty due to the complexity of samples from the distribution. Among them, only samples with confident predictions are used jointly with labeled samples to train the classification head.
• Figure 2: (a) AUROC ($\%$) and (b) Error rate ($\%$) compared to OpenMatch on Mini-ImageNet. (c) The effect of $M$ using in $\text{M}_\text{Inference}$ during inference. We use the same fold to conduct three runs and report the mean of three runs at the top of each bar.
• Figure 3: Graphical model representation of $\mathcal{L}^{\text{N-EDL}}$.
• Figure 4: Case study. U. and O. represent uncertainty and order, respectively. The uncertainty of our method can better distinguish between inliers and outliers than that of OpenMatch. (Samples with blue and yellow backgrounds indicate inliers and outliers, respectively.)