Research map / Learning theory and algorithms / Prediction and decision losses
Understanding Deep Contrastive Learning via Coordinate-wise Optimization
Yuandong Tian
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 |
|---|---|---|---|
| backprop through alpha(theta) instead of coordinate-wise optimization (ablation) | alpha-CL with InfoNCE alpha · CIFAR-10 500 epochs | top-1 accuracy: 87.86 -> 87.48 | Hurt |
| directly set alpha from pairwise distances (alpha-CL-direct) vs InfoNCE | ResNet18 · CIFAR-10 100 epochs | top-1 accuracy: 84.06 -> 85.11 | Improved |
| alpha-CL-direct vs InfoNCE | ResNet18 · STL-10 100 epochs | top-1 accuracy: 78.46 -> 79.32 | Improved |
| alpha-CL-direct vs InfoNCE | ResNet18 · CIFAR-100 100 epochs | top-1 accuracy: 55.70 -> 57.63 | Improved |
| alpha-CL-direct vs InfoNCE | ResNet50 · CIFAR-100 100 epochs | top-1 accuracy: 60.16 -> 62.93 | Improved |
| alpha-CL-direct vs InfoNCE at 500 epochs | ResNet18 · CIFAR-10 | top-1 accuracy: 87.86 -> 88.09 | No clear effect |
| alpha-CL-direct vs InfoNCE at 500 epochs | ResNet50 · STL-10 | top-1 accuracy: 87.90 -> 87.85 | No clear effect |
| replace InfoNCE regularizer with r_H | alpha-CL · CIFAR-10 500 epochs | top-1 accuracy: 87.86 -> 87.92 | No clear effect |
| use squared regularizer r_s instead of InfoNCE (ablation) | alpha-CL · STL-10 500 epochs | top-1 accuracy: 83.70 -> 81.65 | Hurt |
| use quadratic loss instead of InfoNCE | contrastive loss · CIFAR-10 500 epochs | top-1 accuracy: 87.86 -> 73.58 | Hurt |
| use r_gamma regularizer instead of InfoNCE | alpha-CL · STL-10 500 epochs | top-1 accuracy: 83.70 -> 83.47 | No clear effect |
| switch backbone ResNet18 to ResNet50 | InfoNCE · CIFAR-10 500 epochs | top-1 accuracy: 87.86 -> 90.19 | Improved |
Closest papers by meaning
- Generalization Analysis for Deep Contrastive Representation Learning (AAAI 2025 · similarity 0.89)
- Dissecting Supervised Contrastive Learning (ICML 2021 · similarity 0.88)
- On the Similarities of Embeddings in Contrastive Learning (ICML 2025 · similarity 0.87)
- Your contrastive learning problem is secretly a distribution alignment problem (NeurIPS 2024 · similarity 0.87)
- Self-Supervised Contrastive Learning is Approximately Supervised Contrastive Learning (NeurIPS 2025 · similarity 0.87)
- When Can We Approximate Wide Contrastive Models with Neural Tangent Kernels and Principal Component Analysis? (AAAI 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.
