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

Learning with Structural Labels for Learning with Noisy Labels

CVPR 2024 · Label-efficient and robust learning › Noisy labels · approach: Sample selection and co-training · 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 Learning with Structural Labels (LSL), which estimates soft structural labels via reverse k-NN over feature space and trains an LNL model with them on top of SSR, improving robustness to instance-dependent and real-world label noise.

Problem

Learning with noisy labels (LNL) methods rely heavily on the model's own predictions for relabeling and sample selection, which degrades generalization and yields complex feature manifolds under noise.

Approach

Estimate class-conditional data distributions with a reverse k-NN graph over normalized features to produce soft structural labels for all samples (not only selected ones), and train with them via strong augmentation and mixup, added on top of the SSR framework.

Key results

• On CIFAR-10 IDN at noise 0.20, LSL reaches 97.13% vs 96.68% for InstanceGM and 96.49% for SSR.
• On CIFAR-10N-Worst, LSL achieves 94.57% vs 93.50% for reproduced SSR and 92.56% for DivideMix.
• On Animal-10N, LSL reaches 89.1% vs 88.5% for SSR; on WebVision top-1 LSL achieves 81.40% vs 80.92% for SSR.

Measured effects reported in the paper

Surprising findings

• On Red Mini-ImageNet (CNWL), InstanceGM outperforms LSL at all noise rates, so LSL is only second-best there.
• On CIFAR-10 IDN at noise 0.50, LSL is slightly below DivideMix-level baseline (95.81 vs 95.90 for InstanceGM).

Limitations

• Structural labels depend on noisy model predictions and feature quality; kst must be tuned, since too small or too large values degrade performance.
• Gains over strongest baselines are often small (under 1 point) and single-run results are reported.
• Results on Red Mini-ImageNet and CIFAR-10 IDN at 0.50 noise are not best.

Experimental setup

Datasets: CIFAR-10 CIFAR-100 CIFAR-10N CIFAR-100N Animal-10N Red Mini-ImageNet (CNWL) Mini-WebVision ImageNet ILSVRC12

Benchmarks: CIFAR-10 IDN CIFAR-100 IDN CIFAR-10N CIFAR-100N Red Mini-ImageNet (CNWL) Animal-10N WebVision ILSVRC2012

Metrics: test accuracy (%)

Baselines: CE Mixup DivideMix SSR InstanceGM CAL ELR SOP+ MentorMix FaMUS TCL Sel-CL+

Builds on: SSR DivideMix mixup InstanceGM

Closest papers by meaning

• Joint Optimization Framework for Learning With Noisy Labels (CVPR 2018 · similarity 0.92)
• Understanding and Utilizing Deep Neural Networks Trained with Noisy Labels (ICML 2019 · similarity 0.90)
• Distilling Effective Supervision From Severe Label Noise (CVPR 2020 · similarity 0.89)
• On the Role of Label Noise in the Feature Learning Process (ICML 2025 · similarity 0.89)
• Directional Label Diffusion Model for Learning from Noisy Labels (CVPR 2025 · similarity 0.89)
• Handling Label Noise via Instance-Level Difficulty Modeling and Dynamic Optimization (NeurIPS 2025 · similarity 0.88)