Research map / Label-efficient and robust learning / Recognition and long-tail
$K$-Shot Contrastive Learning of Visual Features With Multiple Instance Augmentations
Haohang Xu, Hongkai Xiong, Guo-Jun Qi
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 pairwise cosine similarity with projection onto K=5 augmentation subspace (K-shot contrastive learning) | MoCo v2 (ResNet-50, 200 epochs) · ImageNet linear eval | top-1 accuracy: 67.5 -> 68.8 | Improved |
| apply K-shot subspace contrastive loss over longer pretraining | MoCo v2 (rerun, 800 epochs) · ImageNet linear eval | top-1 accuracy: 70.6 -> 71.4 | Improved |
| increase number of augmentations K from 1 to 3 (ablation) | KSCL, rho=40% · ImageNet linear eval | top-1 accuracy: 67.2 -> 68.5 | Improved |
| increase number of augmentations K from 3 to 5 (ablation) | KSCL, rho=40% · ImageNet linear eval | top-1 accuracy: 68.5 -> 68.8 | Improved |
| keep only 40% of eigenvalue mass instead of 90% (ablation) | KSCL, K=5 · ImageNet linear eval | top-1 accuracy: 68.4 -> 68.8 | Improved |
| increase K from 1 to 5 raises training cost (ablation) | KSCL · ImageNet linear eval | training time per epoch (min): 16 -> 37 | Hurt |
| K=1 KSCL (projection length) versus cosine similarity of MoCo v2 (ablation) | KSCL K=1 · ImageNet linear eval | top-1 accuracy: 67.5 -> 67.2 | No clear effect |
| use KSCL pretrained backbone instead of MoCo v2 (200 epochs) for detection | Faster RCNN on VOC · VOC07 test detection | AP: 57.0 -> 57.1 | Improved |
| use KSCL pretrained backbone instead of MoCo v2 (800 epochs) for detection | Faster RCNN on VOC · VOC07 test detection | AP75: 64.0 -> 64.2 | Improved |
| use a MoCo queue-based KSCL compared with SimCLR large batch baseline gain | KSCL vs MoCo v2 over SimCLR 8192 baseline · ImageNet linear eval | top-1 accuracy: 66.6 -> 68.8 | Improved |
Closest papers by meaning
- Intriguing Properties of Contrastive Losses (NeurIPS 2021 · similarity 0.88)
- What Should Not Be Contrastive in Contrastive Learning (ICLR 2021 · similarity 0.87)
- A Statistical Theory of Contrastive Learning via Approximate Sufficient Statistics (NeurIPS 2025 · similarity 0.85)
- OrCo: Towards Better Generalization via Orthogonality and Contrast for Few-Shot Class-Incremental Learning (CVPR 2024 · similarity 0.85)
- $\mathbb{X}$-Sample Contrastive Loss: Improving Contrastive Learning with Sample Similarity Graphs (ICLR 2025 · similarity 0.85)
- Variational Supervised Contrastive Learning (NeurIPS 2025 · similarity 0.84)
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.
