Research map / Label-efficient and robust learning / Recognition and long-tail
Ranking Info Noise Contrastive Estimation: Boosting Contrastive Learning via Ranked Positives
David T. Hoffmann, Nadine Behrmann, Juergen Gall, Thomas Brox, Mehdi Noroozi
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 ranked class and superclass positives with RINCE instead of SCL-out | Cifar-100 representation learning · Cifar-100 fine | accuracy: 76.70 ± 0.29 -> 77.59 ± 0.21 | Improved |
| Use ranked positives with RINCE instead of SCL | Cifar-100 representation learning · Cifar-100 fine | R@1: 74.45 ± 0.39 -> 77.47 ± 0.16 | Improved |
| Use ranked positives with RINCE instead of SCL-out | ImageNet-100 representation learning · ImageNet-100 | accuracy: 84.18 -> 84.90 | Improved |
| Rank frame, shot, and video clips instead of using InfoNCE frame positives only | Unsupervised video representation learning · HMDB | Top 1 accuracy: 41.5 -> 44.9 | Improved |
| Rank frame, shot, and video clips instead of using InfoNCE frame positives only | Unsupervised video representation learning · UCF | Retrieval mAP: 0.0688 -> 0.1395 | Improved |
| Treat distant video clips as true positives and force all positives to be similar using log-out (ablation) | Unsupervised video representation learning · HMDB | Top 1 accuracy: 41.5 -> 41.4 | Hurt |
| Treat distant video clips as hard negatives rather than positives (ablation) | Unsupervised video representation learning · UCF | Top 1 accuracy: 74.5 -> 75.2 | Improved |
| Use log-in rather than log-out when positives may include noisy distant clips (ablation) | Unsupervised video representation learning · HMDB | Top 1 accuracy: 41.4 -> 42.6 | Improved |
| Set the rank-2 similarity threshold to 0.45 rather than 0.35 (ablation) | RINCE-out-in on ImageNet-100 · ImageNet-100 OOD | AUROC: 60.92 -> 62.92 | Improved |
| Use memory bank size 4096 rather than 2048 (ablation) | RINCE-out-in on Cifar-100 · Cifar-100 linear evaluation | accuracy: 77.03 -> 77.46 | Improved |
| Set rank-2 temperature to 0.225 rather than 0.125 (ablation) | RINCE-in on Cifar-100 · Cifar-100 linear evaluation | accuracy: 75.98 -> 77.18 | Improved |
| Train for 2000 rather than 1000 epochs (ablation) | Supervised contrastive learning and RINCE · Cifar-100 | accuracy | No clear effect |
Closest papers by meaning
- Robust Contrastive Learning Against Noisy Views (CVPR 2022 · similarity 0.86)
- Contrastive Predictive Coding Done Right for Mutual Information Estimation (ICLR 2026 · similarity 0.86)
- When Softmax Fails at the Top: Extreme‑Value Corrections for InfoNCE (ICML 2026 · similarity 0.86)
- Embedding Contrastive Unsupervised Features to Cluster in- and Out-of-Distribution Noise in Corrupted Image Datasets (ECCV 2022 · similarity 0.85)
- On the Similarities of Embeddings in Contrastive Learning (ICML 2025 · similarity 0.84)
- Mitigating the Impact of False Negative in Dense Retrieval with Contrastive Confidence Regularization (AAAI 2024 · 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.
