Research map / Label-efficient and robust learning / Self-supervised and contrastive
When Does Contrastive Visual Representation Learning Work?
Check 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 |
|---|---|---|---|
| pretrain SimCLR on images from the downstream domain rather than another domain | SimCLR ResNet-50 · iNat21, ImageNet, Places365, and GLC20 | top-1 accuracy on iNat21, ImageNet, and Places365; MAP on GLC20: 0.493 vs. 0.373 on iNat21; 0.644 vs. 0.519 on ImageNet; 0.501 vs. 0.486 on Places365; 0.769 vs. 0.716 on GLC20 | Helped |
| reduce in-domain pretraining data from 1M to 500k images (ablation) | SimCLR ResNet-50 · iNat21, ImageNet, and Places365 | linear evaluation accuracy: iNat21: 0.493 -> 0.477; ImageNet: 0.644 -> 0.629; Places365: 0.501 -> 0.499 | Hurt |
| reduce in-domain pretraining data from 1M to 250k images (ablation) | SimCLR ResNet-50 · iNat21, ImageNet, and Places365 | linear evaluation accuracy: iNat21: 0.493 -> 0.451; ImageNet: 0.644 -> 0.608; Places365: 0.501 -> 0.485 | Hurt |
| pool pretraining images from different domains instead of using an equally sized in-domain dataset (ablation) | SimCLR ResNet-50 · iNat21, ImageNet, and Places365 | linear evaluation accuracy: iNat21 pooled 0.444 vs. in-domain 0.477; ImageNet 0.597 vs. 0.629; Places365 0.467 vs. 0.499 | Hurt |
| combine two self-supervised representations by concatenating their features (ablation) | SimCLR representations pretrained on ImageNet and iNat21 · ImageNet and iNat21 | linear evaluation accuracy: ImageNet: 0.647 -> 0.641; iNat21: 0.506 -> 0.520 | Mixed |
| add supervised iNat21 features to ImageNet SimCLR features (ablation) | ImageNet SimCLR representation · iNat21 | linear evaluation accuracy: +4.7% | Helped |
| add supervised iNat21 features to ImageNet SimCLR features (ablation) | ImageNet SimCLR representation · ImageNet | linear evaluation accuracy: -4.2% | Hurt |
| resize ImageNet images before SimCLR pretraining (ablation) | SimCLR ResNet-50 · ImageNet | linear evaluation accuracy: 0.647 -> 0.632 | Hurt |
| increase the number of ImageNet classes represented in 100k pretraining images from 200 to 1000 (ablation) | SimCLR ResNet-50 · iNat21 | linear evaluation accuracy: 0.314 -> 0.290 | Hurt |
| increase the number of ImageNet classes represented in 100k pretraining images from 200 to 1000 (ablation) | SimCLR ResNet-50 · Places365 | linear evaluation accuracy: 0.390 -> 0.429 | Helped |
| pretrain SimCLR on downsampled images (ablation) | SimCLR ResNet-50 · ImageNet | linear evaluation accuracy: around 15% for Downsample (2x); around 34% for Downsample (4x) | Hurt |
| use more fine-grained evaluation labels (ablation) | SimCLR representations versus supervised training · iNat21, ImageNet, and Places365 | top-1 accuracy: null | Hurt |
Closest papers by meaning
- A Simple Framework for Contrastive Learning of Visual Representations (ICML 2020 · similarity 0.92)
- On the duality between contrastive and non-contrastive self-supervised learning (ICLR 2023 · similarity 0.92)
- Supervised Contrastive Learning (NeurIPS 2020 · similarity 0.91)
- Boost Supervised Pretraining for Visual Transfer Learning: Implications of Self-Supervised Contrastive Representation Learning (AAAI 2022 · similarity 0.91)
- Revisiting Contrastive Methods for Unsupervised Learning of Visual Representations (NeurIPS 2021 · similarity 0.91)
- A Theoretical Analysis of Self-Supervised Learning for Vision Transformers (ICLR 2025 · similarity 0.88)
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.