Evidential Deep Active Learning for Semi-Supervised Classification

TL;DR

This work tackles label-scarce semi-supervised classification by introducing EDALSSC, which explicitly models prediction uncertainty for both labeled and unlabeled data. It combines evidential deep learning for labeled samples with a T-conorm-based aggregation of ignorance and conflict for unlabeled ones, and introduces a dynamic Dirichlet scaling mechanism to balance class count and evidence. A mid-to-late training sample-selection strategy selects high-uncertainty unlabeled samples, guided by a learning objective that blends supervised evidence-based loss with unsupervised consistency and uncertainty terms. Across CIFAR-10/100, SVHN, and Fashion-MNIST, EDALSSC consistently outperforms strong baselines and ablations, demonstrating improved sample efficiency and more reliable uncertainty estimates for active learning in semi-supervised settings.

Abstract

Semi-supervised classification based on active learning has made significant progress, but the existing methods often ignore the uncertainty estimation (or reliability) of the prediction results during the learning process, which makes it questionable whether the selected samples can effectively update the model. Hence, this paper proposes an evidential deep active learning approach for semi-supervised classification (EDALSSC). EDALSSC builds a semi-supervised learning framework to simultaneously quantify the uncertainty estimation of labeled and unlabeled data during the learning process. The uncertainty estimation of the former is associated with evidential deep learning, while that of the latter is modeled by combining ignorance information and conflict information of the evidence from the perspective of the T-conorm operator. Furthermore, this article constructs a heuristic method to dynamically balance the influence of evidence and the number of classes on uncertainty estimation to ensure that it does not produce counter-intuitive results in EDALSSC. For the sample selection strategy, EDALSSC selects the sample with the greatest uncertainty estimation that is calculated in the form of a sum when the training loss increases in the latter half of the learning process. Experimental results demonstrate that EDALSSC outperforms existing semi-supervised and supervised active learning approaches on image classification datasets.

Figures (4)

• Figure 1: Entropy of the prediction result misclassify the same samples in CIFAR-10 and CIFAR-100. Blue: Shannon entropy for SSCAL. Red: Ignorance, Conflict and Uncertainty estimation for the proposed method.
• Figure 2: Overview of EDALSSC. It consists of two modules: (a) Image Task Model trains the base model jointly with labeled and unlabeled data, optimizing a composite objective that combines evidence-based cross entropy loss on labeled samples, uncertainty estimation (using T-conorm to comprehensively consider ignorance information and conflict information) on unlabeled samples, and consistency loss across all samples; (b) Uncertainty Estimation Selector selects supplementary samples with higher uncertainty estimates, which have higher total uncertainty estimation during the middle-to-late stages of training, when training loss increases.
• Figure 3: Performance comparison of image classification on four benchmark datasets.
• Figure 4: The t-SNE visualization of the CIFAR10, SVHN and Fashion-MNIST dataset, highlighting the sample selection behavior of EDALSSC. Selected samples are marked in black, while unlabeled samples are shown in color. The boundaries of the class are marked with red lines.