Research map / Label-efficient and robust learning / Self-supervised and contrastive
Improving Contrastive Learning by Visualizing Feature Transformation
Rui Zhu, Bingchen Zhao, Jingen Liu, Zhenglong Sun, Chang Wen 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 |
|---|---|---|---|
| add positive extrapolation (alpha_ex=2.0) (ablation) | MoCoV1 · ImageNet-100 (200 epoch, linear readout) | top-1 accuracy: 71.1 -> 72.8 | Improved |
| replace positive extrapolation with positive interpolation (alpha_ex=2) (ablation) | MoCoV1 · ImageNet-100 (200 epoch, linear readout) | top-1 accuracy: 71.1 -> 67.4 | Hurt |
| add negative interpolation (alpha_in=1.6) (ablation) | MoCoV1 · ImageNet-100 (200 epoch, linear readout) | top-1 accuracy: 71.1 -> 74.6 | Improved |
| combine positive extrapolation and negative interpolation (ablation) | MoCoV1 · ImageNet-100 (200 epoch, linear readout) | top-1 accuracy: 71.10 -> 76.87 | Improved |
| combine positive extrapolation and negative interpolation (ablation) | MoCoV2 · ImageNet-100 (200 epoch, linear readout) | top-1 accuracy: 75.61 -> 78.33 | Improved |
| use dimension-level mixing instead of feature-level mixing (ablation) | MoCoV1 + both FT · ImageNet-100 (200 epoch, linear readout) | top-1 accuracy: 76.87 -> 77.21 | Improved |
| double the queue by adding interpolated negatives to original queue (2K) (ablation) | MoCoV1 + neg FT · ImageNet-100 (200 epoch, linear readout) | top-1 accuracy: 74.64 -> 74.73 | No clear effect |
| double original queue size without feature transformation (K to 2K) (ablation) | MoCoV1 · ImageNet-100 (200 epoch, linear readout) | top-1 accuracy: 71.10 -> 71.40 | No clear effect |
| start feature transformation at epoch 0 instead of never (ResNet-50) (ablation) | MoCoV1 + both FT · ImageNet-100 (ResNet-50, FT start epoch) | top-1 accuracy: 71.1 -> 76.9 | Improved |
| start feature transformation at epoch 80 instead of never (ResNet-50) (ablation) | MoCoV1 + both FT · ImageNet-100 (ResNet-50, FT start epoch) | top-1 accuracy: 71.1 -> 72.2 | Improved |
| train 500 epochs with FT instead of 200 (ablation) | MoCoV2 + both FT · ImageNet-100 (200 epoch, linear readout) | top-1 accuracy: 80.7 -> 81.5 | Mixed |
| add positive extrapolation (ablation) | BYOL · ImageNet-1K (100 epoch) | top-1 accuracy: 66.5 -> 67.2 | Improved |
Closest papers by meaning
- Contrastive Learning with Synthetic Positives (ECCV 2024 · similarity 0.89)
- Decoupled Contrastive Learning (ECCV 2022 · similarity 0.89)
- What Should Not Be Contrastive in Contrastive Learning (ICLR 2021 · similarity 0.89)
- Conditional Negative Sampling for Contrastive Learning of Visual Representations (ICLR 2021 · similarity 0.88)
- MetAug: Contrastive Learning via Meta Feature Augmentation (ICML 2022 · similarity 0.88)
- Linear Differential Vision Transformer: Learning Visual Contrasts via Pairwise Differentials (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.
