Research map / Label-efficient and robust learning / Self-supervised and contrastive
CaCo: Both Positive and Negative Samples are Directly Learnable via Cooperative-Adversarial Contrastive Learning
Xiao Wang, Yuhang Huang, Dan Zeng, 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 |
|---|---|---|---|
| take positives from the shared memory bank in addition to negatives (ablation) | CaCo (batch 1024, 200 epochs) · ImageNet-1k linear eval | top-1 accuracy: 70.6 -> 71.3 | Improved |
| take negatives from the adversarially trained memory bank in addition to bank positives (ablation) | CaCo (batch 1024, 200 epochs) · ImageNet-1k linear eval | top-1 accuracy: 70.7 -> 71.3 | Improved |
| use a memory bank for positives only vs none (ablation) | CaCo (batch 1024, 200 epochs) · ImageNet-1k linear eval | top-1 accuracy: 70.2 -> 70.7 | Improved |
| use a memory bank for negatives only vs none (ablation) | CaCo (batch 1024, 200 epochs) · ImageNet-1k linear eval | top-1 accuracy: 70.2 -> 70.6 | Improved |
| increase memory bank from 4M to 12M parameters (37268 to 98304 negatives) (ablation) | CaCo (batch 4096, 200 epochs) · ImageNet-1k linear eval | top-1 accuracy: 71.7 -> 72.1 | Improved |
| increase batch size from 256 to 4096 (ablation) | CaCo (200 epochs) · ImageNet-1k linear eval | top-1 accuracy: 70.9 -> 72.0 | Improved |
| replace AdCo adversarial-only memory bank with cooperative-adversarial bank | AdCo · ImageNet-1k linear eval (200 epochs, batch 256) | top-1 accuracy: 68.2 -> 70.9 | Improved |
| train longer, 200 to 800 epochs at batch 4096 | CaCo · ImageNet-1k linear eval | top-1 accuracy: 72.0 -> 75.3 | Improved |
| add multi-crop augmentation, batch 2048 | CaCo (800 epochs) · ImageNet-1k linear eval | top-1 accuracy: 75.3 -> 75.7 | Improved |
| compare CaCo to NNCLR under single crop at batch 4096 | NNCLR · ImageNet-1k linear eval (800 epochs) | top-1 accuracy: 74.9 -> 75.3 | Hurt |
| use CaCo features instead of NNCLR for transfer | NNCLR · Transfer to CIFAR100 | linear accuracy: 79.0 -> 76.0 | Mixed |
| use CaCo features instead of MoCo v2 for transfer | MoCo v2 · Transfer to Aircraft | linear accuracy: 50.9 -> 66.0 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- AdCo — AdCo: Adversarial Contrast for Efficient Learning of Unsupervised Representations From Self-Trained Negative Adversaries (CVPR 2021) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- AdCo: Adversarial Contrast for Efficient Learning of Unsupervised Representations From Self-Trained Negative Adversaries (CVPR 2021 · similarity 0.93)
- Robust Contrastive Learning Using Negative Samples with Diminished Semantics (NeurIPS 2021 · similarity 0.89)
- Conditional Negative Sampling for Contrastive Learning of Visual Representations (ICLR 2021 · similarity 0.89)
- Relative Contrastive Loss for Unsupervised Representation Learning (ECCV 2022 · similarity 0.88)
- On the Similarities of Embeddings in Contrastive Learning (ICML 2025 · similarity 0.87)
- $\mathbb{X}$-Sample Contrastive Loss: Improving Contrastive Learning with Sample Similarity Graphs (ICLR 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.
