Research map / Label-efficient and robust learning / Self-supervised and contrastive
Optimal Positive Generation via Latent Transformation for Contrastive Learning
Yinqi Li, Hong Chang, Bingpeng Ma, Shiguang Shan, Xilin CHEN
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 |
|---|---|---|---|
| replace Gaussian latent transformation with learned optimal positives | synthetic-image contrastive learning · ImageNet-1K linear eval | top-1 (%): 48.73 -> 53.25 | Improved |
| replace steering with learned instance-specific positives | synthetic-image contrastive learning · ImageNet-1K linear eval | top-1 (%): 51.19 -> 53.25 | Improved |
| train on BigBiGAN synthetic images with COP-Gen instead of real images with SimCLR | contrastive learning · ImageNet-1K linear eval | top-1 (%): 49.44 -> 53.25 | Improved |
| remove data augmentation Tx (use only Tz) (ablation) | COP-Gen · ImageNet-100 linear eval | top-1 (%): 63.96 -> 37.40 | Hurt |
| remove color augmentation (ablation) | COP-Gen · ImageNet-100 linear eval | top-1 (%): 63.96 -> 55.26 | Hurt |
| remove crop augmentation (ablation) | COP-Gen · ImageNet-100 linear eval | top-1 (%): 63.96 -> 43.84 | Hurt |
| remove all data space augmentation (ablation) | COP-Gen · ImageNet-100 linear eval | top-1 (%): 63.96 -> 36.16 | Hurt |
| increase navigator training samples from 200K to 400K (ablation) | COP-Gen · ImageNet-100 linear eval | top-1 (%): 63.96 -> 64.12 | No clear effect |
| increase navigator training samples from 200K to 600K (ablation) | COP-Gen · ImageNet-100 linear eval | top-1 (%): 63.96 -> 63.56 | No clear effect |
| use linear latent trajectory instead of nonlinear MLP Tz (ablation) | COP-Gen · ImageNet-100 linear eval | top-1 (%): 63.96 -> 61.18 | Hurt |
| learned optimal positives vs steering without data augmentation | COP-Gen · ImageNet-100 linear eval | top-1 (%): 28.68 -> 37.40 | Improved |
| COP-Gen vs real-image NNCLR on VOC detection | COP-Gen · PASCAL VOC | AP: 51.70 -> 53.09 | Improved |
Closest papers by meaning
- Contrastive Learning with Synthetic Positives (ECCV 2024 · similarity 0.89)
- Improving Contrastive Learning by Visualizing Feature Transformation (ICCV 2021 · similarity 0.88)
- Contrastive Learning Can Find An Optimal Basis For Approximately View-Invariant Functions (ICLR 2023 · similarity 0.88)
- Contrastive Learning Inverts the Data Generating Process (ICML 2021 · similarity 0.88)
- 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 0.86)
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.
