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

LCGC: Learning from Consistency Gradient Conflicting for Class-Imbalanced Semi-Supervised Debiasing

AAAI 2025 · Label-efficient and robust learning › Semi-supervised learning · approach: Confidence-based pseudo-label selection · 2 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 LCGC, a debiasing scheme for class-imbalanced semi-supervised learning that trains an over-biased classifier by up-weighting pseudo-label gradients that conflict with black-image-refined logits, then subtracts black-image logits at test time.

Problem

Classifiers trained on class-imbalanced data under semi-supervised learning become biased toward majority classes, degrading accuracy on balanced test sets.

Approach

Theoretically motivates a black baseline image as logit reference via integrated gradients, then modifies the consistency gradient by projecting it along the debiased-direction gradient when they conflict (angle >90°), deliberately producing an over-biased classifier that is then corrected by subtracting baseline logits at inference.

Key results

• On CIFAR-10-LT with gamma_l=gamma_u=100, FixMatch+LCGC reaches 84.9 bACC vs 83.6 for CDMAD; ReMixMatch+LCGC reaches 85.7 vs 85.5 for CDMAD.
• With unknown unlabeled distribution, CIFAR-10-LT gamma_l=100, gamma_u=150: FixMatch+LCGC 84.0 bACC, about 1.7 points above the second-best.
• On CIFAR-100-LT (gamma=20), FixMatch+LCGC achieves 55.3 bACC, 1.0 point above the second-best (CDMAD 54.3).

Measured effects reported in the paper

Surprising findings

• Solid black and white baseline images perform similarly, and other solid colors also work, attributed to BatchNorm standardizing near-uniform features; noise baselines perform worse.
• Gains from LCGC grow with the imbalance ratio, per the authors' conclusion.

Limitations

• Theoretical analysis relies on strong assumptions (e.g., interpolating classifier, gradient flow) that are not verified in practice.
• Gains over CDMAD are often small (e.g., under 1 point in several settings) relative to reported standard errors.

Experimental setup

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

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

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

Baselines: FixMatch ReMixMatch DARP cRT LDAM-DRW CReST ABC CoSSL SAW CDMAD DebiasPL UDAL

Builds on: FixMatch ReMixMatch CDMAD Integrated Gradients DARP

Closest papers by meaning

• CDMAD: Class-Distribution-Mismatch-Aware Debiasing for Class-Imbalanced Semi-Supervised Learning (CVPR 2024 · similarity 0.92)
• Debiased Learning From Naturally Imbalanced Pseudo-Labels (CVPR 2022 · similarity 0.89)
• ABC: Auxiliary Balanced Classifier for Class-imbalanced Semi-supervised Learning (NeurIPS 2021 · similarity 0.89)
• Twice Class Bias Correction for Imbalanced Semi-supervised Learning (AAAI 2024 · similarity 0.87)
• BaCon: Boosting Imbalanced Semi-supervised Learning via Balanced Feature-Level Contrastive Learning (AAAI 2024 · similarity 0.86)
• Semi-supervised Multi-label Learning with Balanced Binary Angular Margin Loss (NeurIPS 2024 · similarity 0.86)