Research map / Label-efficient and robust learning / Self-supervised and contrastive
Decoupled Contrastive Learning
Chun-Hsiao Yeh, Cheng-Yao Hong, Yen-Chi Hsu, Tyng-Luh Liu, Yubei Chen, Yann LeCun
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 |
|---|---|---|---|
| remove positive term from InfoNCE denominator (DCL) (ablation) | SimCLR · ImageNet-1K (bs256, 200 ep) | linear top-1: 61.8 -> 65.9 | Improved |
| apply DCL at small batch size 32 (ablation) | SimCLR · ImageNet-1K (bs32) | linear top-1: 56.8 -> 61.5 | Improved |
| apply DCL at batch size 512 (ablation) | SimCLR · ImageNet-1K (bs512) | linear top-1: 64.0 -> 65.8 | Improved |
| apply DCL at batch size 1024 (ablation) | SimCLR · ImageNet-1K (bs1024) | linear top-1: 65.1 -> 66.1 | Improved |
| add DCL to MoCo-v2 with small queue 64 (ablation) | MoCo-v2 · ImageNet-1K | linear top-1: 63.9 -> 65.8 | Improved |
| add DCL to MoCo-v2 with queue 65536 (ablation) | MoCo-v2 · ImageNet-1K | linear top-1: 67.5 -> 67.7 | No clear effect |
| add von Mises-Fisher positive weighting (DCLW) (ablation) | DCL · ImageNet-1K (bs256, 200 ep) | linear top-1: 65.9 -> 66.9 | Improved |
| tune temperature and learning rate (ablation) | DCL · ImageNet-1K | linear top-1: 65.9 -> 67.8 | Improved |
| add BYOL asymmetric augmentation (ablation) | DCL · ImageNet-1K | linear top-1: 67.8 -> 68.2 | Improved |
| add DCL when training only 100 epochs (ablation) | SimCLR · ImageNet-1K (100 ep) | linear top-1: 57.5 -> 64.6 | Improved |
| combine DCL with NNCLR at batch size 512 (ablation) | NNCLR · ImageNet-1K | linear top-1: 71.7 -> 72.3 | Improved |
| use DCL over InfoNCE for temperature sensitivity (ablation) | SimCLR · CIFAR10 temperature sweep | std of kNN accuracy: 1.44 -> 0.98 | Improved |
Closest papers by meaning
- Contrastive Learning with Synthetic Positives (ECCV 2024 · similarity 0.92)
- On the duality between contrastive and non-contrastive self-supervised learning (ICLR 2023 · similarity 0.91)
- Supervised Contrastive Learning (NeurIPS 2020 · similarity 0.91)
- Self-Contrastive Learning: Single-Viewed Supervised Contrastive Framework Using Sub-network (AAAI 2023 · similarity 0.90)
- Self-Supervised Contrastive Learning is Approximately Supervised Contrastive Learning (NeurIPS 2025 · similarity 0.90)
- Bridging Mini-Batch and Asymptotic Analysis in Contrastive Learning: From InfoNCE to Kernel-Based Losses (ICML 2024 · similarity 0.89)
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.
