Diverse Teacher-Students for Deep Safe Semi-Supervised Learning under Class Mismatch

TL;DR

This work tackles safe semi-supervised learning under class distribution mismatch by decoupling seen-class classification from unseen-class detection. It introduces Diverse Teacher-Students (DTS), a dual-teacher framework where an Inlier Teacher-Student pair specializes in seen-class learning and an Outlier Teacher-Student pair focuses on unseen-class detection, guided by a novel soft-weighting uncertainty score that softly separates unseen samples and enables a ($K$+1)-class supervisory signal. The method jointly trains both branches with labeled and unlabeled data, using targeted losses such as Logit Matching and consistency regularization to enhance robustness and utilization of unlabeled data. Empirically, DTS outperforms a wide range of baselines across SVHN, CIFAR-10/100, and STL-10 under varying mismatch ratios, and exhibits improved unseen-class detection (AUROC) alongside stable optimization. The approach requires no dataset-specific priors and demonstrates strong practical appeal for real-world deployment, with code and models publicly available.

Abstract

Semi-supervised learning can significantly boost model performance by leveraging unlabeled data, particularly when labeled data is scarce. However, real-world unlabeled data often contain unseen-class samples, which can hinder the classification of seen classes. To address this issue, mainstream safe SSL methods suggest detecting and discarding unseen-class samples from unlabeled data. Nevertheless, these methods typically employ a single-model strategy to simultaneously tackle both the classification of seen classes and the detection of unseen classes. Our research indicates that such an approach may lead to conflicts during training, resulting in suboptimal model optimization. Inspired by this, we introduce a novel framework named Diverse Teacher-Students (\textbf{DTS}), which uniquely utilizes dual teacher-student models to individually and effectively handle these two tasks. DTS employs a novel uncertainty score to softly separate unseen-class and seen-class data from the unlabeled set, and intelligently creates an additional ($K$+1)-th class supervisory signal for training. By training both teacher-student models with all unlabeled samples, DTS can enhance the classification of seen classes while simultaneously improving the detection of unseen classes. Comprehensive experiments demonstrate that DTS surpasses baseline methods across a variety of datasets and configurations. Our code and models can be publicly accessible on the link https://github.com/Zhanlo/DTS.

Figures (9)

• Figure 1: Illustration of the safe semi-supervised learning setting with class distribution mismatch. Take the categories of CIFAR-10 dataset as an example, in which the animal categories are seen classes and the other categories are unseen classes.
• Figure 2: A common safe semi-supervised learning framework. Such frameworks generally train a singular model to undertake dual tasks: classification of seen-classes and detection of unseen-classes.
• Figure 3: Overview of DTS. DTS offers a robust approach to deep, safe semi-supervised learning for datasets with unseen classes. It integrates four pivotal modules, starting with a pre-trained teacher model to establish the Inlier Teacher-Student (ITS) and Outlier Teacher-Student (OTS) frameworks. Each iteration advances with the student model refining the teacher model. ITS and OTS then engage a soft-weighting mechanism to capitalize on unlabeled data, achieving a harmonized optimization of the training objectives.
• Figure 4: (a) and (b) respectively depict the AUROC trend for our proposed DTS and SAFE-STUDENT models, as well as the supervised baseline, under varying Labeled/Unlabeled Classes Mismatch ratios on CIFAR-10.
• Figure 5: Ablation Analysis: (a), (b), and (c) depict the impact of various modules on the classification performance of the DTS framework across different values of 'm' and class mismatch ratios, illustrating the contribution of each component under varying conditions.