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

GradTrack: Detecting Noisy Labels via Temporal Trajectories of Class-wise Gradient Misalignment

NeurIPS 2026 · Label-efficient and robust learning › Noisy labels · approach: Noise-aware loss and pairwise methods

Read the paperCheck your idea against 108,967 papers

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

Proposes GradTrack, a noisy-label detection method that uses temporal trajectories of class-wise gradient misalignment rather than scalar loss, confidence, or gradient-magnitude signals.

Problem

Detecting incorrectly labeled training samples so their harmful effect on model learning can be mitigated; existing signals (loss, confidence, gradient magnitude) capture only the scale of errors and ignore gradient geometry and temporal dynamics.

Approach

Models the geometric structure and temporal evolution of class-wise gradient misalignment across training, turning gradient dynamics into a detection signal instead of a scalar error magnitude.

Closest papers by meaning

• Learning Discriminative Dynamics with Label Corruption for Noisy Label Detection (CVPR 2024 · similarity 0.89)
• Revisiting Interpolation for Noisy Label Correction (AAAI 2025 · similarity 0.88)
• Handling Label Noise via Instance-Level Difficulty Modeling and Dynamic Optimization (NeurIPS 2025 · similarity 0.87)
• NoiseGPT: Label Noise Detection and Rectification through Probability Curvature (NeurIPS 2024 · similarity 0.87)
• Estimating Noisy Class Posterior with Part-level Labels for Noisy Label Learning (CVPR 2024 · similarity 0.87)
• Radial-Angular Geometry for Reliable Update Diagnosis in Noisy-Label Learning (NeurIPS 2026 · similarity 0.87)