Research map / Label-efficient and robust learning / Recognition and long-tail
Multi-label Contrastive Predictive Coding
Jiaming Song, Stefano Ermon
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 ML-CPC with J_0.01 instead of CRD | CIFAR-100 distillation, WRN-40-2 to WRN-16-2 · CIFAR-100 | top-1 test accuracy (%): 75.48 -> 75.83 | Improved |
| Use ML-CPC with J_0.01 instead of CRD | CIFAR-100 distillation, resnet110 to resnet20 · CIFAR-100 | top-1 test accuracy (%): 71.46 -> 71.27 | Hurt |
| Use ML-CPC with J_0.01 instead of CRD | CIFAR-100 distillation, resnet32x4 to resnet8x4 · CIFAR-100 | top-1 test accuracy (%): 75.51 -> 75.37 | Hurt |
| Use smaller alpha (J_0.05) rather than alpha 1.0 ML-CPC (ablation) | CIFAR-100 distillation, WRN-40-2 to WRN-40-1 · CIFAR-100 | top-1 test accuracy (%): 74.18 -> 74.27 | Improved |
| Use smaller alpha (J_0.05) rather than alpha 1.0 ML-CPC (ablation) | CIFAR-100 distillation, ResNet50 to MobileNetV2 · CIFAR-100 | top-1 test accuracy (%): 68.80 -> 70.04 | Improved |
| Replace CPC with ML-CPC using a high-to-low alpha curriculum | Unsupervised representation learning · CIFAR-10, 200 epochs | top-1 accuracy: 83.28 -> 86.16 | Improved |
| Use the J_10.0 to J_0.1 curriculum rather than fixed alpha 1.0 ML-CPC (ablation) | Unsupervised representation learning · CIFAR-100, 200 epochs | top-1 accuracy: 61.80 -> 64.05 | Improved |
| Use the J_10.0 to J_0.1 curriculum rather than fixed alpha 1.0 ML-CPC (ablation) | Unsupervised representation learning · CIFAR-100, 1000 epochs | top-1 accuracy: 70.85 -> 70.03 | Hurt |
| Replace CPC with ML-CPC using a high-to-low alpha curriculum | Unsupervised representation learning · ImageNet, 30 epochs | Top1: 43.45 -> 43.86 | Improved |
| Replace CPC with ML-CPC using a high-to-low alpha curriculum | Unsupervised representation learning · ImageNet, 30 epochs | Top5: 67.42 -> 67.67 | Improved |
| Use ML-CPC instead of CPC | Mutual information estimation with 128 positive and negative samples · Correlated Gaussians, 20 dimensions | computation time for 200 updates (seconds): 1.15 ± 0.06 -> 1.14 ± 0.04 | No clear effect |
| Reduce alpha for ML-CPC (ablation) | Mutual information estimation · Correlated Gaussians, 20 dimensions | estimation bias | Improved |
Closest papers by meaning
- Contrastive Predictive Coding Done Right for Mutual Information Estimation (ICLR 2026 · similarity 0.87)
- Connecting Jensen–Shannon and Kullback–Leibler Divergences: A New Bound for Representation Learning (NeurIPS 2025 · similarity 0.87)
- Tight Mutual Information Estimation With Contrastive Fenchel-Legendre Optimization (NeurIPS 2022 · similarity 0.87)
- CLUB: A Contrastive Log-ratio Upper Bound of Mutual Information (ICML 2020 · similarity 0.86)
- Contrastive Learning Can Find An Optimal Basis For Approximately View-Invariant Functions (ICLR 2023 · similarity 0.86)
- Variational Supervised Contrastive Learning (NeurIPS 2025 · similarity 0.85)
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.
