Research map / Label-efficient and robust learning / Self-supervised and contrastive
A Simple Framework for Contrastive Learning of Visual 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 composed random cropping and color distortion rather than a single augmentation (ablation) | SimCLR · ImageNet | linear evaluation top-1 accuracy | Helped |
| Increase color distortion strength from 1/8 to 1 (ablation) | SimCLR · ImageNet, 100 epochs | linear evaluation top-1 accuracy: 59.6 -> 63.2 | Helped |
| Add Gaussian blur (ablation) | SimCLR · ImageNet, 100 epochs | linear evaluation top-1 accuracy: 63.2 -> 64.5 | Helped |
| Add random horizontal flipping (ablation) | SimCLR · ImageNet, 100 epochs | linear evaluation top-1 accuracy: 63.4 -> 64.5 | Helped |
| Use stronger color distortion (ablation) | Supervised ResNet-50 · ImageNet | top-1 accuracy: 77.0 -> 75.7 | Hurt |
| Add a learnable nonlinear projection head (ablation) | SimCLR · ImageNet | linear evaluation accuracy: +3% | Helped |
| Use a nonlinear projection head instead of no projection (ablation) | SimCLR · ImageNet | linear evaluation accuracy: >10% | Helped |
| Replace NT-Xent with NT-Logistic and semi-hard negative mining (ablation) | SimCLR · ImageNet | linear evaluation top-1 accuracy: 63.9 -> 57.9 | Hurt |
| Use ℓ2 normalization and a tuned temperature instead of unnormalized embeddings (ablation) | SimCLR · ImageNet | linear evaluation top-1 accuracy: 57.2 -> 64.4 | Helped |
| Use square-root rather than linear learning-rate scaling (ablation) | SimCLR with LARS · ImageNet, batch size 256, 100 epochs | linear evaluation top-1 accuracy: 57.5 -> 62.8 | Helped |
| Increase batch size when training for a fixed small number of epochs (ablation) | SimCLR · ImageNet | linear evaluation accuracy | Helped |
| Train for more epochs (ablation) | SimCLR · ImageNet | linear evaluation accuracy | Helped |
Closest papers by meaning
- When Does Contrastive Visual Representation Learning Work? (CVPR 2022 · similarity 0.92)
- On the duality between contrastive and non-contrastive self-supervised learning (ICLR 2023 · similarity 0.92)
- Self-Contrastive Learning: Single-Viewed Supervised Contrastive Framework Using Sub-network (AAAI 2023 · similarity 0.92)
- BECLR: Batch Enhanced Contrastive Few-Shot Learning (ICLR 2024 · similarity 0.89)
- $\mathbb{X}$-Sample Contrastive Loss: Improving Contrastive Learning with Sample Similarity Graphs (ICLR 2025 · similarity 0.89)
- Self-Supervised Contrastive Learning is Approximately Supervised Contrastive Learning (NeurIPS 2025 · similarity 0.88)
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.