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

Radial-Angular Geometry for Reliable Update Diagnosis in Noisy-Label Learning

NeurIPS 2026 · Label-efficient and robust learning › Noisy labels · approach: Noise-aware loss and pairwise methods · 0 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 Relative Geometric Conflict (RGC), which scores noisy-label reliability by comparing the classifier-layer observed-label gradient with an EMA-teacher reference gradient, fused with a last-layer Fisher-trace magnitude term to produce trust weights.

Problem

Noisy-label methods estimate sample reliability from forward-space signals (loss, confidence, entropy) that measure difficulty rather than whether the observed label induces a trustworthy parameter update; hard clean and mislabeled samples overlap in these signals.

Approach

Reframes reliability as diagnosis of the observed-label update: uses the sample-wise empirical Fisher trace (last-layer closed form) for update magnitude and an angular conflict term between the observed-label gradient and an EMA-teacher reference gradient for direction, fused into a continuous trust weight.

Key results

• RGC reaches 93.5/92.4/86.9/87.6 on CIFAR-10 (sym 20/50/80, asym 40) and 90.1 average, the top entry in each CIFAR-10 column versus IDO's 85.6 average.
• On the loss-overlapped diagnostic, loss achieves 0.503 AUROC while trace reaches 0.823 AUROC/0.801 AUPRC.
• Ablation on CIFAR-10 20% symmetric: Trace 91.5, Teacher 91.2, Trace+Teacher 92.6, full RGC 93.5.

Surprising findings

• At the extreme 20% data ratio, Trace slightly surpasses RGC in accuracy (84.7 vs 84.3) and HC-FP (0.46 vs 0.49).
• Teacher-only supervision is less stable than trace-based variants in some settings, suggesting teacher supervision alone can drift.

Limitations

• Relies on the EMA reference being more stable than the observed-label update; fails under teacher drift toward corrupted labels or extreme noise.
• Last-layer trace and conflict are classifier-only proxies, giving sufficient conditions rather than general guarantees.
• Beta fusion interval and EMA hyperparameters are sufficient conditions rather than optimal tuning; gains shrink at low data ratios.

Experimental setup

Datasets: CIFAR-10 CIFAR-100 Animal-10N Food-101N

Benchmarks: CIFAR-10 CIFAR-100 Animal-10N Food-101N

Metrics: test accuracy AUROC AUPRC HC-FP HCPR NoisyRec

Baselines: Standard CE GCE DivideMix RRL UNICON ProMix LongReMix L2B PSSCL DCD EarlyCut IDO

Builds on: IDO DivideMix mean teacher

Closest papers by meaning

• Label Noise: Ignorance Is Bliss (NeurIPS 2024 · similarity 0.85)
• Learning from Noisy Labels via Conditional Distributionally Robust Optimization (NeurIPS 2024 · similarity 0.84)
• Revisiting Interpolation for Noisy Label Correction (AAAI 2025 · similarity 0.84)
• NoiseGPT: Label Noise Detection and Rectification through Probability Curvature (NeurIPS 2024 · similarity 0.84)
• Graph–Smoothed Bayesian Black-Box Shift Estimator and Its Information Geometry (NeurIPS 2025 · similarity 0.83)
• Collaborative Refining for Learning from Inaccurate Labels (NeurIPS 2024 · similarity 0.83)