Research map / Label-efficient and robust learning / Semi-supervised learning

CDMAD: Class-Distribution-Mismatch-Aware Debiasing for Class-Imbalanced Semi-Supervised Learning

CVPR 2024 · Label-efficient and robust learning › Semi-supervised learning · approach: Confidence-based pseudo-label selection · 20 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes CDMAD, which subtracts the classifier's logits on a solid-color image from logits for unlabeled training samples and test samples to debias pseudo-labels and predictions in class-imbalanced semi-supervised learning under class distribution mismatch.

Problem

Class-imbalanced semi-supervised learning where classifiers are biased toward majority classes and pseudo-labels inherit this bias, especially when the unlabeled set's class distribution is unknown and differs from the labeled set's.

Approach

Measure the classifier's per-class bias by its logits on a solid-color image (assumed uninformative about labels) at each training iteration, and subtract this bias from logits when refining pseudo-labels and test predictions; this acts as a post-hoc logit adjustment whose prior is implicitly estimated from the model rather than from label counts.

Key results

• On CIFAR-10-LT with gamma_l=gamma_u=100, FixMatch+CDMAD reaches 83.6 bACC vs 83.2 for the strongest baseline CoSSL (FixMatch+CoSSL) and 71.5 for FixMatch.
• With gamma_l=100 and unknown gamma_u=1 on CIFAR-10-LT, ReMixMatch+CDMAD achieves 89.9 bACC, outperforming ReMixMatch*+DARP+cRT at 88.7 (ReMixMatch* uses estimated unlabeled distribution).
• On Small-ImageNet-127 (32x32), FixMatch+CDMAD reaches 48.4 bACC vs 46.9 for FixMatch+ABC.

Measured effects reported in the paper

Surprising findings

• Replacing the solid-color image with random-pixel noise (uniform, Bernoulli, normal) degrades ReMixMatch+CDMAD, while different solid colors perform similarly.
• Post-hoc logit adjustment (LA) can hurt performance under unknown, mismatched unlabeled distributions (e.g., ReMixMatch*+SAW+LA drops to 74.2 bACC at gamma_u=1).

Limitations

• The solid-color image assumption to be uninformative for labels lacks a firm theoretical basis, as acknowledged by the authors.
• Fisher consistency proof holds only in the population setting, not directly for practical training.
• Gains under class-distribution mismatch depend on the base SSL algorithm; ReMixMatch needs a supervised loss modification and no distribution alignment.

Experimental setup

Datasets: CIFAR-10-LT CIFAR-100-LT STL-10-LT Small-ImageNet-127

Benchmarks: CIFAR-10-LT CIFAR-100-LT STL-10-LT Small-ImageNet-127

Metrics: balanced accuracy (bACC) geometric mean (GM)

Baselines: FixMatch ReMixMatch DARP CReST ABC CoSSL SAW DASO UDAL L2AC ACR FreeMatch

Builds on: FixMatch ReMixMatch logit adjustment cRT DARP

Closest papers by meaning

• LCGC: Learning from Consistency Gradient Conflicting for Class-Imbalanced Semi-Supervised Debiasing (AAAI 2025 · similarity 0.92)
• Debiased Learning From Naturally Imbalanced Pseudo-Labels (CVPR 2022 · similarity 0.91)
• Distribution Aligning Refinery of Pseudo-label for Imbalanced Semi-supervised Learning (NeurIPS 2020 · similarity 0.90)
• Unsupervised Learning for Class Distribution Mismatch (ICML 2025 · similarity 0.89)
• Twice Class Bias Correction for Imbalanced Semi-supervised Learning (AAAI 2024 · similarity 0.89)
• Semi-supervised Multi-label Learning with Balanced Binary Angular Margin Loss (NeurIPS 2024 · similarity 0.87)