Research map / Label-efficient and robust learning / Self-supervised and contrastive
Learning Weakly-supervised Contrastive Representations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use attributes-determined clusters with Cl-InfoNCE instead of instance-ID clusters (SimCLR) (ablation) | ResNet-50 visual representation learning · Wider Attribute | Top-1 Acc.: 40.2 -> 45.5 | Helped |
| Use downstream-label clusters with Cl-InfoNCE instead of attributes-determined clusters (ablation) | ResNet-50 visual representation learning · Wider Attribute | Top-1 Acc.: 45.5 -> 49.9 | Helped |
| Replace Cl-InfoNCE with cross-entropy cluster prediction (ablation) | Attributes-determined clusters · UT-Zappos50K | Top-1 Acc.: 84.6 over 82.7 | Hurt |
| Replace Cl-InfoNCE with cross-entropy cluster prediction (ablation) | Attributes-determined clusters · Wider Attribute | Top-1 Acc.: 45.5 over 39.4 | Hurt |
| Replace Cl-InfoNCE with cross-entropy cluster prediction (ablation) | Attributes-determined clusters · CUB-200-2011 | Top-1 Acc.: 20.6 over 17.5 | Hurt |
| Use CMC instead of attributes-determined clusters with Cl-InfoNCE (ablation) | Weakly-supervised representation learning · CUB-200-2011 | Top-1 Acc.: 32.7 over 20.6 | Helped |
| Use attributes-determined clusters with Cl-InfoNCE instead of CMC (ablation) | Weakly-supervised representation learning · Wider Attribute | Top-1 Acc.: 45.5 over 34.1 | Helped |
| Use attributes-determined clusters with Cl-InfoNCE instead of CMC (ablation) | Weakly-supervised representation learning · UT-Zappos50K | Top-1 Acc.: 84.6 over 83.7 | Helped |
| Use attributes-determined clusters with Cl-InfoNCE instead of CMC (ablation) | Weakly-supervised representation learning · CUB-200-2011 | Top-1 Acc.: 20.6 over 32.7 | Hurt |
| Increase the number of selected high-entropy attributes (ablation) | Attributes-determined clusters + Cl-InfoNCE · UT-Zappos50K | Downstream performance | Mixed |
| Construct clusters using K-means and train with Cl-InfoNCE instead of PCL (ablation) | Unsupervised representation learning · ImageNet-100 | Linear evaluation | Helped |
| Add an InfoNCE objective with an additional head to Cl-InfoNCE + Kmeans (ablation) | ImageNet-100 representation learning · ImageNet-100 | Top-1 Accuracy (%): 82.2 $\pm$ 0.4 -> 82.6 $\pm$ 0.2 | Helped |
Closest papers by meaning
- Weakly-Supervised Contrastive Learning for Imprecise Class Labels (ICML 2025 · similarity 0.88)
- Rethinking Weak Supervision in Helping Contrastive Learning (ICML 2023 · similarity 0.87)
- A Theory-Driven Self-Labeling Refinement Method for Contrastive Representation Learning (NeurIPS 2021 · similarity 0.85)
- I-Con: A Unifying Framework for Representation Learning (ICLR 2025 · similarity 0.85)
- Contrastive Learning is Spectral Clustering on Similarity Graph (ICLR 2024 · similarity 0.84)
- Breaking the Batch Barrier (B3) of Contrastive Learning via Smart Batch Mining (NeurIPS 2025 · similarity 0.84)
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.