Research map / Label-efficient and robust learning / Self-supervised and contrastive
Contrastive Learning with Synthetic Positives
Dewen Zeng, Xinrong Hu, Yawen Wu, Xiaowei Xu, Yiyu Shi
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| add synthetic hard positives from diffusion feature interpolation to SimCLR | SimCLR · CIFAR10 (linear) | top-1 accuracy: 91.47 -> 94.37 | Improved |
| add synthetic hard positives to SimCLR | SimCLR · CIFAR100 (linear) | top-1 accuracy: 65.78 -> 72.01 | Improved |
| add synthetic positive to SimCLR on ImageNet100 | SimCLR ResNet-50 · ImageNet100 | top-1 accuracy: 75.75 -> 79.62 | Improved |
| replace augmented view with synthetic positive (no naive augmentation pair) (ablation) | CLSP-SimCLR · CIFAR100 (linear) | top-1 accuracy: 72.01 -> 54.46 | Hurt |
| replace augmented view with synthetic positive (no naive augmentation pair) (ablation) | CLSP-SimCLR · CIFAR10 (linear) | top-1 accuracy: 91.47 -> 92.86 | Improved |
| use synthetic positive instead of an extra naively augmented positive (ablation) | CLSP-SimCLR · CIFAR100 (linear) | top-1 accuracy: 68.30 -> 72.01 | Improved |
| increase positive candidate set size from 1 to 8 (ablation) | CLSP-SimCLR · CIFAR100 (linear) | top-1 accuracy: 65.84 -> 72.01 | Improved |
| increase candidate set size beyond 8 to 32 (ablation) | CLSP-SimCLR · CIFAR100 (linear) | top-1 accuracy: 72.01 -> 72.04 | No clear effect |
| increase loss weight lambda beyond 1.0 (ablation) | CLSP-SimCLR · CIFAR100 (linear) | top-1 accuracy: 72.01 -> 69.07 | Hurt |
| use larger feature interpolation weight w (ablation) | CLSP-SimCLR · CIFAR10/CIFAR100 | linear accuracy | Hurt |
| use more than one additional synthetic positive per anchor (ablation) | CLSP-SimCLR · CIFAR10 | top-1 accuracy: 94.37 -> 94.35 | No clear effect |
| use feature interpolation instead of RCG embedding-conditioned generation (ablation) | CLSP-SimCLR · CIFAR100 | top-1 accuracy: 67.90 -> 72.01 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- All4One — All4One: Symbiotic Neighbour Contrastive Learning via Self-Attention and Redundancy Reduction (ICCV 2023) · better on 6 of 8 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Decoupled Contrastive Learning (ECCV 2022 · similarity 0.92)
- Synthetic Data Can Also Teach: Synthesizing Effective Data for Unsupervised Visual Representation Learning (AAAI 2023 · similarity 0.90)
- Improving Contrastive Learning by Visualizing Feature Transformation (ICCV 2021 · similarity 0.89)
- Self-Supervised Contrastive Learning is Approximately Supervised Contrastive Learning (NeurIPS 2025 · similarity 0.87)
- On the Similarities of Embeddings in Contrastive Learning (ICML 2025 · similarity 0.87)
- A Statistical Theory of Contrastive Learning via Approximate Sufficient Statistics (NeurIPS 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.
