Research map / Label-efficient and robust learning / Self-supervised and contrastive
Towards Domain-Agnostic Contrastive Learning
Vikas Verma, Thang Luong, Kenji Kawaguchi, Hieu Pham, Quoc Le
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 |
|---|---|---|---|
| Replace Gaussian noise with Mixup noise | DACL contrastive pretraining · Fashion-MNIST (permuted, tabular) | accuracy: 75.8 -> 81.4 | Improved |
| Add geometric and binary Mixup noise alongside linear Mixup (ablation) | DACL · Fashion-MNIST and CIFAR10 (permuted, tabular) | accuracy: 81.4 -> 82.4; 37.6 -> 39.7 | Improved |
| Replace Gaussian noise with Mixup noise | DACL contrastive pretraining · CIFAR-10 and CIFAR-100 | accuracy: 56.1 -> 81.3; 29.8 -> 46.5 | Improved |
| Add geometric and binary Mixup noise alongside linear Mixup (ablation) | DACL · CIFAR-10 and CIFAR-100 | accuracy: 81.3 -> 83.8; 46.5 -> 52.7 | Improved |
| Combine DACL loss with SimCLR loss | SimCLR · CIFAR-10, CIFAR-100, and ImageNet | accuracy: 93.4 -> 94.3; 73.8 -> 75.5; 73.4 -> 74.4 (Top 1) | Improved |
| Use DACL instead of image-specific augmentation-based contrastive learning | ImageNet representation learning · ImageNet | Top 1 accuracy: 73.4 -> 24.6 | Hurt |
| Use Mixup noise instead of Gaussian noise | DACL contrastive pretraining · ImageNet | Top 1 accuracy: 10.2 -> 24.6 | Improved |
| Apply DACL rather than InfoGraph | Graph representation learning with GIN · MUTAG | accuracy: 86.74 ± 1.28 -> 85.31 ± 1.34 | Mixed |
| Apply DACL rather than InfoGraph | Graph representation learning with GIN · PTC-MR, REDDIT-BINARY, IMDB-BINARY, and IMDB-MULTI | accuracy: 57.09 ± 1.52 -> 59.24± 2.57; 63.52± 1.66 -> 66.92 ± 3.38; 63.97 ± 2.05 -> 64.71±2.13; 39.28 ± 1.43 -> 40.16 ± 1.50 | No clear effect |
| Apply DACL rather than InfoGraph | Graph representation learning with GIN · REDDIT-M5K | accuracy: 42.89±0.62 -> 42.86 ± 1.11 | No clear effect |
| Use DACL representation learning instead of full-network supervised training | 12-layer fully connected network · Fashion-MNIST and CIFAR10 (permuted, tabular) | accuracy: 79.1 -> 81.4; 35.2 -> 37.6 | Improved |
Closest papers by meaning
- $i$-Mix: A Domain-Agnostic Strategy for Contrastive Representation Learning (ICLR 2021 · similarity 0.89)
- A Simple Framework for Contrastive Learning of Visual Representations (ICML 2020 · similarity 0.89)
- Decoupled Contrastive Learning (ECCV 2022 · similarity 0.88)
- On the duality between contrastive and non-contrastive self-supervised learning (ICLR 2023 · similarity 0.88)
- DomCLP: Domain-wise Contrastive Learning with Prototype Mixup for Unsupervised Domain Generalization (AAAI 2025 · similarity 0.88)
- Self-Supervised Contrastive Learning is Approximately Supervised Contrastive Learning (NeurIPS 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.
