Research map / Label-efficient and robust learning / Self-supervised and contrastive
Weakly-Supervised Contrastive Learning for Imprecise Class Labels
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 a graph-theoretic weakly-supervised contrastive loss (WSC) that replaces discrete positive/negative pairs with continuous semantic similarity built from weak labels, and plugs it into noisy-label and partial-label learning methods.
Problem
Supervised contrastive learning requires reliable same-class labels, but real annotations are noisy or ambiguous (noisy labels, partial labels), so positive/negative pairs cannot be trusted.
Approach
Replace discrete positive/negative pairs with a continuous semantic similarity S computed from the weak supervision (via a recovery matrix S satisfying P(y|x)=S(x)P(q|x)), so that the perturbed augmentation graph's edge weights approximate supervised connectivity; the resulting loss is a weighted spectral contrastive objective with regularization terms.
Key results
- On CIFAR-100 with 90% symmetric noise, WSC surpasses the prior best TCL by 6.85 points (61.32 vs 54.50).
- On CIFAR-100 with partial ratio 0.3, WSC outperforms the previous best method by 7.8 points (69.15 vs 61.35 for PiCO).
- WSC achieves best accuracy on most NLL and PLL settings, e.g. 95.22 on CIFAR-10 asymmetric noise 0.4 and 94.41 on CIFAR-10 partial ratio 0.8.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Integrate WSC with the baseline on CIFAR-100 at symmetric noise 0.9 | NLL baseline · CIFAR-100, symmetric noise 0.9 | accuracy: 54.50 -> 61.32; 6.85% | Helped |
| Integrate WSC with the baseline on CIFAR-100 at partial ratio 0.3 | PLL baseline · CIFAR-100, partial ratio 0.3 | accuracy: 61.35 -> 69.15; 7.8% | Helped |
| Integrate WSC with the baseline across the reported noisy-label settings | NLL baseline · CIFAR-10 and CIFAR-100, reported noise settings | accuracy | Mixed |
| Integrate WSC with the baseline across the reported partial-label settings | PLL baseline · CIFAR-10, CIFAR-100, and CUB-200, reported partial ratios | accuracy | Helped |
| Use weakly-supervised contrastive learning rather than self-supervised connectivity alone | WSC representation learning · Theoretical linear-probe error bound | linear probe error bound | Helped |
Limitations
- Theoretical bounds rely on finite-vertex and uniform class distribution assumptions
- Requires estimating the transition matrix or recovery matrix S, whose quality bounds performance
- Experiments mostly on simulated noise and partial labels with CIFAR-scale datasets
Experimental setup
Datasets: CIFAR-10 CIFAR-100 CUB-200 CIFAR-10N Clothing1M CIFAR-100-H
Benchmarks: CIFAR-10 CIFAR-100 CUB-200 CIFAR-10N Clothing1M
Metrics: classification accuracy
Baselines: CE DivideMix ELR SOP GFWS ULAREF ProMix MOIT Sel-CL TCL LWS PRODEN
Builds on: supervised contrastive learning spectral contrastive loss PiCO TCL
Closest papers by meaning
- Rethinking Weak Supervision in Helping Contrastive Learning (ICML 2023 · similarity 0.91)
- A Theory-Driven Self-Labeling Refinement Method for Contrastive Representation Learning (NeurIPS 2021 · similarity 0.90)
- Self-Supervised Contrastive Learning is Approximately Supervised Contrastive Learning (NeurIPS 2025 · similarity 0.88)
- Debiased Contrastive Learning (NeurIPS 2020 · similarity 0.88)
- On the Similarities of Embeddings in Contrastive Learning (ICML 2025 · similarity 0.86)
- Continuous Contrastive Learning for Long-Tailed Semi-Supervised Recognition (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.