Research map / Label-efficient and robust learning / Self-supervised and contrastive
AdCo: Adversarial Contrast for Efficient Learning of Unsupervised Representations From Self-Trained Negative Adversaries
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Train negative adversaries by maximizing the contrastive loss instead of passively queuing negatives | AdCo · ImageNet linear evaluation, 200 epochs | Top-1 accuracy: 68.6 vs. 67.5 | Helped |
| Use multi-crop augmentations with trainable negative adversaries | AdCo · ImageNet linear evaluation, 200 epochs | Top-1 accuracy: 73.2 vs. 72.7 | Helped |
| Use multi-crop augmentations with trainable negative adversaries | AdCo · ImageNet linear evaluation, 800 epochs | Top-1 accuracy: 75.7 vs. 75.3 | Helped |
| Symmetrize the contrastive loss (ablation) | AdCo · ImageNet linear evaluation, 200 epochs, single-crop | Top-1 accuracy: 68.6 -> 70.6 | Helped |
| Symmetrize the loss with multi-crop augmentations (ablation) | AdCo* · ImageNet linear evaluation, 200 epochs, multi-crop | Top-1 accuracy: 73.2 -> 73.6 | Helped |
| Reduce negative adversaries from 65,536 to 16,384 (ablation) | AdCo · ImageNet linear evaluation, 200 epochs, single-crop | Top-1 accuracy: 68.6 -> 68.6 | No clear effect |
| Reduce negative adversaries from 65,536 to 8,192 (ablation) | AdCo · ImageNet linear evaluation, 200 epochs, single-crop | Top-1 accuracy: 68.6 -> 68.4 | No clear effect |
| Reduce negative adversaries from 65,536 to 16,384 while using symmetric loss (ablation) | AdCo · ImageNet linear evaluation, 200 epochs, single-crop | Top-1 accuracy: 70.6 -> 70.2 | No clear effect |
| Reduce negative adversaries from 65,536 to 8,192 while using symmetric loss (ablation) | AdCo · ImageNet linear evaluation, 200 epochs, single-crop | Top-1 accuracy: 70.6 -> 70.2 | No clear effect |
| Use the alternative adversarial loss based on p(q_i|n_k) rather than maximizing the contrastive loss directly (ablation) | AdCo negative adversary updates · ImageNet pretraining | Downstream performance | Hurt |
| Use AdCo rather than MoCo v2 for pretraining | ResNet-50 · ImageNet linear evaluation, 10 epochs | Top-1 accuracy: 5% | Helped |
| Use AdCo rather than MoCo v2 for pretraining | ResNet-50 · ImageNet linear evaluation, 800 epochs | Top-1 accuracy: 72.8% vs. 71.1% | Helped |
Closest papers by meaning
- Conditional Negative Sampling for Contrastive Learning of Visual Representations (ICLR 2021 · similarity 0.88)
- Robust Contrastive Learning Using Negative Samples with Diminished Semantics (NeurIPS 2021 · similarity 0.88)
- Contrastive Self-Supervised Learning Leads to Higher Adversarial Susceptibility (AAAI 2023 · similarity 0.88)
- Efficient Adversarial Contrastive Learning via Robustness-Aware Coreset Selection (NeurIPS 2023 · similarity 0.88)
- CAT: Contrastive Adversarial Training for Evaluating the Robustness of Protective Perturbations in Latent Diffusion Models (ICML 2025 · similarity 0.87)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.86)
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.