Research map / Label-efficient and robust learning / Self-supervised and contrastive
Synthetic Data Can Also Teach: Synthesizing Effective Data for Unsupervised Visual Representation Learning
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 |
|---|---|---|---|
| Jointly generate hard samples and hard positive pairs instead of using SimCLR alone | SimCLR · CIFAR-10 linear classification | top-1 accuracy: 90.37 -> 92.94 | Helped |
| Jointly generate hard samples and hard positive pairs instead of using SimCLR alone | SimCLR · CIFAR-100 linear classification | top-1 accuracy: 63.93 -> 67.41 | Helped |
| Jointly generate hard samples and hard positive pairs instead of using SimCLR alone | SimCLR · Fashion-MNIST linear classification | top-1 accuracy: 92.35 -> 93.94 | Helped |
| Jointly generate hard samples and hard positive pairs instead of using SimCLR alone | SimCLR · ImageNet-100 linear classification | top-1 accuracy: 67.45 -> 71.40 | Helped |
| Jointly generate hard samples and hard positive pairs instead of using SimCLR alone | SimCLR · ImageNet-10 linear classification | top-1 accuracy: 83.20 -> 87.40 | Helped |
| Add a standalone generator without joint optimization (ablation) | SimCLR · CIFAR-10 linear classification | top-1 accuracy | Hurt |
| Jointly optimize the generator and contrastive learner to generate hard samples (ablation) | SimCLR · CIFAR-10 linear classification | top-1 accuracy: +0.39% | Helped |
| Generate positive pairs with the generator and its slowly evolving version (ablation) | SimCLR · CIFAR-10 linear classification | top-1 accuracy: +0.62% | Helped |
| Combine jointly generated hard samples with hard positive pairs (ablation) | SimCLR · CIFAR-10 linear classification | top-1 accuracy: 92.94% vs. 90.37% | Helped |
| Enable all proposed components rather than using a separate generator (ablation) | SimCLR · CIFAR-10 with 20% training data | top-1 accuracy: +4.92% | Helped |
| Use a separate generator without the proposed joint-learning components (ablation) | SimCLR · CIFAR-10 with 20% training data | top-1 accuracy: +0.15% | Helped |
Closest papers by meaning
- Generative Models as a Data Source for Multiview Representation Learning (ICLR 2022 · similarity 0.88)
- Do Generated Data Always Help Contrastive Learning? (ICLR 2024 · similarity 0.88)
- Can Generative Models Improve Self-Supervised Representation Learning? (AAAI 2025 · similarity 0.88)
- Personalized Representation from Personalized Generation (ICLR 2025 · similarity 0.87)
- Real-Fake: Effective Training Data Synthesis Through Distribution Matching (ICLR 2024 · similarity 0.86)
- Return of Unconditional Generation: A Self-supervised Representation Generation Method (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.