Research map / Label-efficient and robust learning / Self-supervised and contrastive
Data-Efficient Image Recognition with Contrastive Predictive Coding
Olivier Henaff
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 the CPC v1 implementation with the CPC v2 architecture and training protocol | CPC on ImageNet linear classification · ImageNet | Top-1 accuracy: 48.7% to 71.5% | Improved |
| Increase depth and width to the ResNet-161 encoder (ablation) | CPC v2 · ImageNet linear classification | Top-1 accuracy: +5% | Improved |
| Increase patch size from 64×64 to 80×80 pixels (ablation) | CPC v2 · ImageNet linear classification | Top-1 accuracy: +2% | Improved |
| Use layer normalization instead of batch normalization (ablation) | CPC v2 · ImageNet linear classification | accuracy: +2% | Improved |
| Add bottom-up spatial predictions (ablation) | CPC v2 · ImageNet linear classification | accuracy: +2% | Improved |
| Predict in all four spatial directions (ablation) | CPC v2 · ImageNet linear classification | accuracy: +2.5% | Improved |
| Randomly drop two color channels in each patch (ablation) | CPC v2 · ImageNet linear classification | accuracy: +3% | Improved |
| Add generic transformations, elastic deformations, and color transforms to patch augmentation (ablation) | CPC v2 · ImageNet linear classification | accuracy: +4.5% | Improved |
| Train the classifier on CPC features rather than raw pixels | ImageNet classifiers with limited labels · ImageNet, 1% labels | Top-5 accuracy: 44.1 -> 78.3 | Improved |
| Fine-tune CPC features for the supervised classification task (ablation) | CPC v2 classifier · ImageNet, 1% labels | Top-5 accuracy: 77.1 -> 78.3 | Improved |
| Pre-train the representation with CPC v2 rather than supervised ImageNet pre-training | Faster R-CNN object detector · PASCAL VOC 2007 | mAP: 74.7 -> 76.6 | Improved |
Closest papers by meaning
- Unifying Visual Contrastive Learning for Object Recognition from a Graph Perspective (ECCV 2022 · similarity 0.88)
- A Broad Study on the Transferability of Visual Representations With Contrastive Learning (ICCV 2021 · similarity 0.87)
- A Simple Framework for Contrastive Learning of Visual Representations (ICML 2020 · similarity 0.87)
- Supervised Contrastive Learning (NeurIPS 2020 · similarity 0.87)
- Self-Contrastive Learning: Single-Viewed Supervised Contrastive Framework Using Sub-network (AAAI 2023 · similarity 0.86)
- I-Con: A Unifying Framework for Representation Learning (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.
