Research map / Label-efficient and robust learning / Recognition and long-tail
Which Features are Learnt by Contrastive Learning? On the Role of Simplicity Bias in Class Collapse and Feature Suppression
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 |
|---|---|---|---|
| Combine supervised and unsupervised contrastive losses | Contrastive learning · CIFAR-100 | subclass accuracy: 26.11 -> 41.37 | Helped |
| Combine supervised and unsupervised contrastive losses | Contrastive learning · MNIST RandBit | class accuracy: 61.21 -> 79.37 | Helped |
| Combine supervised and unsupervised contrastive losses | Contrastive learning · CIFAR-100 RandBit | subclass accuracy: 28.13 -> 35.72; 34.11 -> 35.72 | Helped |
| Combine supervised and unsupervised contrastive losses | Contrastive learning · CIFAR-100 RandBit | class accuracy: 61.10 -> 63.94; 52.32 -> 63.94 | Helped |
| Increase embedding width from w=4 to w=64 (ablation) | ResNet18 on CIFAR-10 RandBit · CIFAR-10 RandBit | subclass accuracy: 34.38 -> 71.96 | Helped |
| Increase embedding width from w=4 to w=64 (ablation) | ResNet18 on CIFAR-10 RandBit · CIFAR-10 RandBit | class accuracy: 86.73 -> 96.82 | Helped |
| Increase embedding width from w=64 to w=128 (ablation) | ResNet18 on CIFAR-10 RandBit · CIFAR-10 RandBit | subclass accuracy: 71.96 -> 76.69 | Helped |
| Increase embedding width from w=64 to w=128 (ablation) | ResNet18 on CIFAR-10 RandBit · CIFAR-10 RandBit | class accuracy: 96.82 -> 97.65 | Helped |
| Increase embedding width from w=4 to w=64 (ablation) | ResNet18 on CIFAR-100 RandBit · CIFAR-100 RandBit | subclass accuracy: 11.67 -> 34.11 | Helped |
| Increase embedding width from w=4 to w=64 (ablation) | ResNet18 on CIFAR-100 RandBit · CIFAR-100 RandBit | class accuracy: 23.53 -> 52.32 | Helped |
| Increase embedding width from w=64 to w=128 (ablation) | ResNet18 on CIFAR-100 RandBit · CIFAR-100 RandBit | subclass accuracy: 34.11 -> 38.51 | Helped |
| Increase embedding width from w=64 to w=128 (ablation) | ResNet18 on CIFAR-100 RandBit · CIFAR-100 RandBit | class accuracy: 52.32 -> 57.40 | Helped |
Closest papers by meaning
- Self-Supervised Contrastive Learning is Approximately Supervised Contrastive Learning (NeurIPS 2025 · similarity 0.88)
- An Investigation of Representation and Allocation Harms in Contrastive Learning (ICLR 2024 · similarity 0.88)
- A Statistical Theory of Contrastive Learning via Approximate Sufficient Statistics (NeurIPS 2025 · similarity 0.87)
- On the Similarities of Embeddings in Contrastive Learning (ICML 2025 · similarity 0.87)
- A Tale of Two Classes: Adapting Supervised Contrastive Learning to Binary Imbalanced Datasets (CVPR 2025 · similarity 0.87)
- Generalization Analysis for Supervised Contrastive Representation Learning under Non-IID Settings (ICML 2025 · similarity 0.87)
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.