Research map / Label-efficient and robust learning / Noisy labels

Enhanced Sample Selection with Confidence Tracking: Identifying Correctly Labeled Yet Hard-to-Learn Samples in Noisy Data

AAAI 2025 · Label-efficient and robust learning › Noisy labels · approach: Noise-aware loss and pairwise methods · 5 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 Confidence Tracking (CT), a plug-and-play sample selector that uses Mann-Kendall trend tests on confidence gaps between annotated and other labels to recover correctly labeled hard-to-learn samples under noisy labels, and combines it with existing selectors (GMM, FINE, AUM, DIST, etc.).

Problem

Small-loss sample selection for learning with noisy labels trades precision against recall, since correctly labeled hard-to-learn samples can have high loss similar to mislabeled ones early in training.

Approach

Instead of thresholding loss, track whether the confidence gap between the annotated label and every other class increases over training (via the Mann-Kendall trend test); samples whose gaps all trend upward are treated as correctly labeled and unioned with the baseline selector's picks.

Key results

• GMM+CT reaches 78.46% average accuracy vs 76.85% for GMM on CIFAR-10/100 benchmarks (Table 1).
• DIST+CT improves average sample-selection F1 from 89.12 to 92.15 on CIFAR-100 noisy settings (Table 2).
• CT improves CORES, DivideMix, f-DivideMix and DISC on WebVision and Food-101N, e.g. DivideMix+CT 78.12 vs 77.44 Top1 on WebVision.

Measured effects reported in the paper

Surprising findings

• On Clothing1M the authors found the test set contains noisy labels, so they excluded it as a benchmark (DivideMix 74.12% vs DivideMix+CT 74.18%).

Limitations

• Relies on the early learning phenomenon; too few or too many warm-up epochs harm sample selection.
• Adds a significance level hyperparameter alpha, and gains are smaller under symmetric noise where small-loss already works well.
• Theoretical understanding of early learning is limited; analysis is empirical and in high-dimensional linear setting only.

Experimental setup

Datasets: CIFAR-10 CIFAR-100 CIFAR-10N CIFAR-100N WebVision Food-101N ILSVRC12 Food-101

Benchmarks: CIFAR-10 CIFAR-100 CIFAR-10N CIFAR-100N WebVision ILSVRC12 Food-101N

Metrics: test accuracy precision recall F1-score

Baselines: CE L2D Co-teaching CNLCU HMW GMM FINE AUM DIST CORES DivideMix f-DivideMix

Builds on: Co-teaching AUM DIST FINE DivideMix

Closest papers by meaning

• Enhancing Sample Selection Against Label Noise by Cutting Mislabeled Easy Examples (NeurIPS 2025 · similarity 0.91)
• Regroup Median Loss for Combating Label Noise (AAAI 2024 · similarity 0.87)
• Learning from Noisy Labels via Conditional Distributionally Robust Optimization (NeurIPS 2024 · similarity 0.87)
• Revisiting Interpolation for Noisy Label Correction (AAAI 2025 · similarity 0.87)
• Enhancing Semi-Supervised Learning via Representative and Diverse Sample Selection (NeurIPS 2024 · similarity 0.87)
• Estimation of the Label-Noise Transition Matrix with Performance Guarantees via Selective Classification (NeurIPS 2026 · similarity 0.87)