Intrinsic logit-based debiasing for class-imbalanced semi-supervised learning.
TL;DR
Intrinsic Logit-Based Debiasing (ILBD), a robust post-hoc framework that estimates bias directly from task-relevant data without external dependencies and class priors, and effectively rectifies the classifier's bias directly from the inherent statistical patterns of the training data.
Abstract
In Class-Imbalanced Semi-Supervised Learning (CISSL), classifiers suffer from severe confirmation bias, particularly when the class distribution of unlabeled set are unknown or mismatched with that of the labeled set. Existing debiasing methods often rely on impractical assumptions for classifier bias estimation, such as known class priors or risky external proxies. To overcome these limitations, we propose Intrinsic Logit-Based Debiasing (ILBD), a robust post-hoc framework that estimates bias directly from task-relevant data without external dependencies and class priors. ILBD constructs a comprehensive bias estimator by decomposing output logits into two distinct components: (1) Label-Free Intrinsic Bias, derived from non-target K-1 logits by masking target-class maximum logits to capture the model's background distributional skew; and (2) Label-Guided Learning Bias, derived from target-class maximum logits to quantify confidence disparities between head and tail classes arising from varying learning difficulties. To ensure estimation reliability, we further employ a class-adaptive threshold to filter noisy pseudo-labels. By subtracting this estimated bias at inference, ILBD effectively rectifies the classifier's bias directly from the inherent statistical patterns of the training data, without the need for target-prior assumptions or external data. Extensive experiments on CIFAR-10/100-LT, STL-10-LT, and Small-ImageNet-127 demonstrate the effectiveness of ILBD. Our code is available at https://github.com/aroid721/ILBD.