Research map / Label-efficient and robust learning / Self-supervised and contrastive
Solving Inefficiency of Self-Supervised Representation Learning
Guangrun Wang, Keze Wang, Guangcong Wang, Philip H.S. Torr, Liang Lin
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 |
|---|---|---|---|
| use rank-5 negative instead of hardest negative (ablation) | truncated triplet loss · ImageNet linear eval (20-epoch ablation) | top-1 accuracy: 28.9 -> 29.5 | Improved |
| use rank-52 negative instead of hardest negative (ablation) | truncated triplet loss · ImageNet linear eval (20-epoch ablation) | top-1 accuracy: 28.9 -> 30.0 | Improved |
| use rank-52 instead of hardest triplet (k=1) (ablation) | truncated triplet loss · over-clustering frequency at epoch 180 | Pr(Omega|B): 0.9948 -> 0.0233 | Improved |
| use k=5 instead of k=1 (ablation) | truncated triplet loss · over-clustering frequency at epoch 180 | Pr(Omega|A): 0.9656 -> 0.2105 | Improved |
| use margin -100 instead of -0.3 (ablation) | truncated triplet loss · ImageNet linear eval (20-epoch ablation) | top-1 accuracy: 28.3 -> 30.0 | Improved |
| use margin -1.2 instead of -0.3 (ablation) | truncated triplet loss · ImageNet linear eval (20-epoch ablation) | top-1 accuracy: 28.3 -> 29.8 | Improved |
| use smoothed rank-k instead of rank-k triplet (ablation) | truncated triplet loss · ImageNet linear eval (200 epochs) | top-1 accuracy: 73.6 -> 73.8 | Improved |
| add multi-crop augmentation (ablation) | truncated triplet loss · ImageNet linear eval (200 epochs) | top-1 accuracy: 73.8 -> 74.1 | Improved |
| remove multi-crop from SwAV | SwAV · ImageNet linear eval (200 epochs) | top-1 accuracy: 72.7 -> 69.1 | Hurt |
| train truncated triplet longer | truncated triplet loss · ImageNet linear eval | top-1 accuracy: 73.6 -> 75.9 | Improved |
| train supervised longer | supervised ResNet-50 · ImageNet | top-1 accuracy: 76.3 -> 78.4 | Improved |
| truncated triplet pretraining versus supervised | supervised ImageNet pretraining · COCO 2017 Mask R-CNN | AP box: 40.0 -> 41.7 | Improved |
Closest papers by meaning
- A theoretical study of inductive biases in contrastive learning (ICLR 2023 · similarity 0.90)
- Divide and Contrast: Self-Supervised Learning From Uncurated Data (ICCV 2021 · similarity 0.90)
- Exploring Balanced Feature Spaces for Representation Learning (ICLR 2021 · similarity 0.89)
- Self-Supervised Contrastive Learning is Approximately Supervised Contrastive Learning (NeurIPS 2025 · similarity 0.88)
- Collapse-Proof Non-Contrastive Self-Supervised Learning (ICML 2025 · similarity 0.88)
- Contrastive Learning is Spectral Clustering on Similarity Graph (ICLR 2024 · similarity 0.88)
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.
