Research map / Label-efficient and robust learning / Self-supervised and contrastive
Seed the Views: Hierarchical Semantic Alignment for Contrastive Representation Learning
Haohang Xu, Xiaopeng Zhang, Hao Li, Lingxi Xie, Wenrui Dai, Hongkai Xiong, Qi Tian
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 one KNN-selected positive sample (ablation) | MoCo v2 · ImageNet linear eval (200 epochs) | top-1 accuracy: 67.5 -> 70.4 | Improved |
| add CutMix mixed sample as extra positive (ablation) | CsMl · ImageNet linear eval (200 epochs) | top-1 accuracy: 70.4 -> 71.6 | Improved |
| add multi-crop augmentation (ablation) | CsMl · ImageNet linear eval (200 epochs) | top-1 accuracy: 71.6 -> 74.6 | Improved |
| replace CutMix with MixUp for the mixed sample (ablation) | CsMl · ImageNet linear eval (200 epochs) | top-1 accuracy: 71.6 -> 70.1 | Hurt |
| add a second KNN positive instead of mixed sample (ablation) | CsMl · ImageNet linear eval (200 epochs) | top-1 accuracy: 70.4 -> 70.2 | No clear effect |
| use identical sample as positive (k=0) instead of KNN neighbors (ablation) | MoCo v2 · ImageNet linear eval (200 epochs) | top-1 accuracy: 67.5 -> 68.6 | Hurt |
| add multi-level pulling at stage 3 and 4 (ablation) | CsMl · ImageNet linear eval stage 3 | top-1 accuracy: 60.7 -> 64.0 | Improved |
| add pulling at stage 2, 3 and 4 versus MoCo v2 (ablation) | CsMl · ImageNet 1% labels fine-tuning | top-1 accuracy: 52.4 -> 54.8 | Improved |
| add multi-level pulling (ablation) | CsMl · COCO detection | AP bb: 39.5 -> 40.3 | Improved |
| apply CsMl to BYOL | BYOL · ImageNet linear eval (300 epochs) | top-1 accuracy: 72.5 -> 75.3 | Improved |
| add data mining with pseudo labels | CsMl · ImageNet 1% labels | top-1 accuracy: 62.2 -> 66.7 | Improved |
| compare to MoCo v2 trained for double epochs (matching samples) (ablation) | CsMl · ImageNet linear eval | top-1 accuracy: 67.5 -> 68.5 | Improved |
Closest papers by meaning
- HCSC: Hierarchical Contrastive Selective Coding (CVPR 2022 · similarity 0.91)
- A Simple Framework for Contrastive Learning of Visual Representations (ICML 2020 · similarity 0.89)
- Adaptive and Multi-scale Affinity Alignment for Hierarchical Contrastive Learning (NeurIPS 2025 · similarity 0.89)
- Can Semantic Labels Assist Self-Supervised Visual Representation Learning? (AAAI 2022 · similarity 0.89)
- $\mathbb{X}$-Sample Contrastive Loss: Improving Contrastive Learning with Sample Similarity Graphs (ICLR 2025 · similarity 0.88)
- DH-Set: Improving Vision-Language Alignment with Diverse and Hybrid Set-Embeddings Learning (CVPR 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.
