Uncertainty-aware Sampling for Long-tailed Semi-supervised Learning

TL;DR

This work tackles the bias toward head classes in long-tailed semi-supervised learning by introducing Uncertainty-Aware Dynamic Threshold Selection (UDTS). UDTS integrates model uncertainty into pseudo-label sampling using Monte Carlo dropout and dynamically adjusts per-class thresholds across training stages, guided by a Bayesian risk framework and an uncertainty-driven loss. The approach yields substantial improvements over FixMatch baselines on CIFAR10-LT, CIFAR100-LT, STL-10-LT, and TissueMNIST, and ablation studies confirm the contribution of uncertainty-guided selection, adaptive thresholds, and uncertainty loss. By enabling more reliable pseudo-labels and adapting to evolving learning states, UDTS offers a practical, plug-in enhancement for robust long-tailed semi-supervised learning in real-world scenarios.

Abstract

For semi-supervised learning with imbalance classes, the long-tailed distribution of data will increase the model prediction bias toward dominant classes, undermining performance on less frequent classes. Existing methods also face challenges in ensuring the selection of sufficiently reliable pseudo-labels for model training and there is a lack of mechanisms to adjust the selection of more reliable pseudo-labels based on different training stages. To mitigate this issue, we introduce uncertainty into the modeling process for pseudo-label sampling, taking into account that the model performance on the tailed classes varies over different training stages. For example, at the early stage of model training, the limited predictive accuracy of model results in a higher rate of uncertain pseudo-labels. To counter this, we propose an Uncertainty-Aware Dynamic Threshold Selection (UDTS) approach. This approach allows the model to perceive the uncertainty of pseudo-labels at different training stages, thereby adaptively adjusting the selection thresholds for different classes. Compared to other methods such as the baseline method FixMatch, UDTS achieves an increase in accuracy of at least approximately 5.26%, 1.75%, 9.96%, and 1.28% on the natural scene image datasets CIFAR10-LT, CIFAR100-LT, STL-10-LT, and the medical image dataset TissueMNIST, respectively. The source code of UDTS is publicly available at: https://github.com/yangk/UDTS.

Figures (7)

• Figure 1: Evolution of uncertainty selection during training. Throughout various training stages, both the uncertainty of unlabeled data and the dynamic uncertainty threshold evolve over time. The proposed method prioritizes selecting images with lower uncertainty, enhancing model performance. In the figure, images outlined by red dashed rectangles indicate instances of high uncertainty that have been incorrectly classified.
• Figure 2: Overall framework of UDTS. ${{\tau }_{t1}}$, ${{\tau }_{t2}}$ and ${{\tau }_{t3}}$ are the dynamic uncertainty threshold in different steps. First, the input long-tailed data is divided into labeled and unlabeled data. A model is trained on the labeled data, and predictions are made on the unlabeled data. Additionally, the Monte Carlo Dropout method is employed to estimate the uncertainty of the predictions. After uncertainty-aware selection, relying on various learning states at different training stages, more reliable and diverse pseudo-labels are chosen using adaptive uncertainty thresholds. Simultaneously, the model is adjusted using uncertainty loss. The selected uncertain data and labeled data are combined and sent to the upper layers for supervised learning. This process is repeated in a loop until convergence.
• Figure 3: Theoretical analysis process of UDTS. The red boxes highlight the primary innovative contributions. Firstly, model uncertainty is computed, followed by estimating the risk of the Bayesian classifier. Subsequently, optimizing this classifier occurs. Dynamic thresholds are then designed, and the selection is based on the computed uncertainty. Finally, a loss function is designed to converge the model.
• Figure 4: Analysis of FixMatch and our method in terms of recall, test accuracy, and t-SNE on CIFAR10-LT. Figure (a) and Figure (b) show that the class indexes of the X-axis are sorted by the class size, with C0 as the head class and C9 as the tail class. Figure (c) and Figure (d) show the t-SNE of FixMatch and UDTS respectively. UDTS gets higher recall and test accuracy compared with FixMatch.
• Figure 5: Visualization presenting the results on the CIFAR-10 dataset under the same experimental setup. From left to right, the categories are arranged from head class to tail class, with the most abundant category being “ car” and the least abundant being “ airplane”.