Research map / Label-efficient and robust learning / Recognition and long-tail
Understanding and Generalizing Contrastive Learning from the Inverse Optimal Transport Perspective
Read the paperCheck 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 |
|---|---|---|---|
| add uniformity penalty on coupling (lambda_p=0.5) to IOT-CL U(a) | IOT-CL U(a) · CIFAR-10/CIFAR-100/SVHN (SimCLR) | linear and k-NN top-1 accuracy: e.g. CIFAR-100 linear 66.34 -> 67.59; SVHN k-NN 75.25 -> 84.15 (vs InfoNCE) | Helped |
| IOT-CL loss under U(a) with penalty replaces InfoNCE | SimCLR framework · ImageNet-100 (ResNet50, 100 epochs) | linear and k-NN top-1 accuracy: linear 67.35 -> 68.32; k-NN 53.50 -> 55.28 | Helped |
| sweep uniformity penalty coefficient lambda_p (ablation) | IOT-CL U(a) · CIFAR-10 / CIFAR-100 / SVHN | linear top-1 accuracy: gains at small-moderate lambda_p; lambda_p=10 lowers CIFAR-100 (66.01 vs 66.34 at 100 epochs) | Mixed |
| tighten constraint relaxation (Sinkhorn K=2 with U(a,b)) without penalty (ablation) | IOT-CL loss · CIFAR-10 (SimCLR, 100 epochs) | linear top-1 accuracy: 91.06 vs 90.61 for U(a); but K=4 gives 90.66 and K=1 gives 90.55 | Mixed |
| relax constraints from U(a,b) to U(1) or U(a) in IOT-CL loss (ablation) | IOT-CL loss · CIFAR-100 (SimCLR) | linear top-1 accuracy: U(1) 67.15, U(a) 66.34, U(a,b) K=1 66.74 at 100 epochs | No clear effect |
| Triplet loss (tau->0 InfoNCE) replaces InfoNCE | SimCLR · CIFAR-10 (100 epochs) | linear top-1 accuracy: 70.97 vs 90.61 for InfoNCE | Hurt |
Closest papers by meaning
- Your contrastive learning problem is secretly a distribution alignment problem (NeurIPS 2024 · similarity 0.90)
- Robust Contrastive Learning Against Noisy Views (CVPR 2022 · similarity 0.88)
- Understanding Contrastive Representation Learning through Alignment and Uniformity on the Hypersphere (ICML 2020 · similarity 0.88)
- On the Similarities of Embeddings in Contrastive Learning (ICML 2025 · similarity 0.87)
- Bridging Mini-Batch and Asymptotic Analysis in Contrastive Learning: From InfoNCE to Kernel-Based Losses (ICML 2024 · similarity 0.86)
- $\mathbb{X}$-Sample Contrastive Loss: Improving Contrastive Learning with Sample Similarity Graphs (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.