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.

What problem does this paper attempt to address?

### What problems does this paper attempt to solve? This paper aims to solve the challenges in open - set semi - supervised learning (Open - set SSL), especially when unlabeled data and test data contain new classes (i.e., outliers or OOD samples) that have not been seen in the labeled data. Specifically, the paper focuses on the following two main problems: 1. **Limitations of existing methods**: - **Insufficient uncertainty quantification**: Most existing methods use binary classifiers for anomaly detection. These classifiers can only provide point - estimated probabilities and cannot distinguish different types of uncertainty (such as epistemic uncertainty and aleatoric uncertainty). - **Poor robustness to complex datasets**: When dealing with multi - class information, softmax - based binary classifiers perform poorly in identifying outliers. 2. **How to effectively utilize the information of unlabeled data**: - Existing methods fail to fully utilize the information contained in unlabeled data to improve the model. Especially when distinguishing outliers from normal values, they fail to effectively use this information for model regularization. ### The method proposed in the paper To solve the above problems, the paper proposes a new framework named Adaptive Negative Evidential Deep Learning (ANEDL). The main contributions of ANEDL include: - **Introducing Evidential Deep Learning (EDL) as an anomaly detector**: By regarding the network output as evidence to parameterize the Dirichlet distribution, different types of uncertainty are quantified, and different uncertainty measures are designed for self - training and inference. - **Proposing an adaptive negative optimization strategy**: For unlabeled datasets containing normal and abnormal values, explicitly adjust the EDL detector to output low evidence values to reduce the influence of uncertain categories, and encourage the model to pay more attention to these uncertain categories through adaptive loss weights. - **Jointly optimizing Softmax and EDL**: Combine the Softmax head for representation learning and classification probability prediction, and at the same time use the EDL head to quantify uncertainty, thereby improving classification performance. ### Experimental results The paper conducted extensive experiments on four public datasets (CIFAR - 10, CIFAR - 100, ImageNet - 30, and Mini - ImageNet), and the results show that ANEDL outperforms the existing state - of - the - art methods in different settings. ### Summary This paper solves the challenges of anomaly detection and classification in open - set semi - supervised learning by introducing Evidential Deep Learning and an adaptive negative optimization strategy, significantly improving the performance and robustness of the model.

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

Open World Classification with Adaptive Negative Samples.

Towards Evidential and Class Separable Open Set Object Detection

A Comprehensive Survey on Evidential Deep Learning and Its Applications

An Empirical Study and Analysis on Open-Set Semi-Supervised Learning

Revisiting Essential and Nonessential Settings of Evidential Deep Learning

Evidence-based uncertainty-aware semi-supervised medical image segmentation

Knowledge from Uncertainty in Evidential Deep Learning

Dirichlet-Based Coarse-to-Fine Example Selection For Open-Set Annotation

Uncertainty Estimation by Density Aware Evidential Deep Learning

SSB: Simple but Strong Baseline for Boosting Performance of Open-Set Semi-Supervised Learning

GCL-OSDA: Uncertainty prediction-based graph collaborative learning for open-set domain adaptation

A Novel Semi-Supervised Classification Method Based on Soft Evidential Label Propagation

Robust Semi-supervised Learning by Wisely Leveraging Open-set Data

TEDL: A Two-stage Evidential Deep Learning Method for Classification Uncertainty Quantification

Towards Open Set Video Anomaly Detection

E3 Outlier: A Self-Supervised Framework for Unsupervised Deep Outlier Detection

E-NER: Evidential Deep Learning for Trustworthy Named Entity Recognition

DSV: An Alignment Validation Loss for Self-supervised Outlier Model Selection

Revisiting Deep Semi-supervised Learning: An Empirical Distribution Alignment Framework and Its Generalization Bound

Bridging the Theoretical Bound and Deep Algorithms for Open Set Domain Adaptation