Research map / Label-efficient and robust learning / Self-supervised and contrastive
Self-Supervised Learning via Maximum Entropy Coding
Xin Liu, Zhongdao Wang, Ya-Li Li, Shengjin Wang
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 |
|---|---|---|---|
| Use maximum entropy coding with view consistency instead of established SSL objectives | ResNet-50 pre-training · ImageNet val, linear probe | top-1 accuracy: 70.6 vs 68.1 | Improved |
| Use maximum entropy coding pre-training instead of the strongest listed SSL baseline | ResNet-50 pre-training · ImageNet val, 1% labels | top-1 accuracy: 55.9 vs 55.0 | Improved |
| Use maximum entropy coding pre-training instead of MoCo v2 | ResNet-50 C4 · COCO detection | AP: +0.5 AP | Improved |
| Reduce pre-training from 100 to 50 epochs (ablation) | MEC · ImageNet val | top-1 accuracy: 70.6 -> 67.5 | Hurt |
| Reduce batch size from 1024 to 128 (ablation) | MEC · ImageNet val | top-1 accuracy: 70.6 -> 69.5 | Hurt |
| Set distortion εd² to 0.02 instead of 0.06 (ablation) | MEC · ImageNet val | top-1 accuracy: 70.6 -> 70.5 | No clear effect |
| Set distortion εd² to 1.0 instead of 0.06 (ablation) | MEC · ImageNet val | top-1 accuracy: 70.6 -> 69.5 | Hurt |
| Remove the momentum encoder (ablation) | MEC · ImageNet val | top-1 accuracy: 70.6 -> 69.0 | Hurt |
| Use first-order instead of fourth-order Taylor expansion (ablation) | MEC · ImageNet val | top-1 accuracy: 70.6 -> 70.0 | Hurt |
| Increase Taylor expansion order from 1 to 4 (ablation) | MEC · COCO detection | AP: +1.5 AP | Improved |
| Use symmetric rather than asymmetric projector networks (ablation) | MEC · ImageNet val | top-1 accuracy: 70.6 -> 70.1 | No clear effect |
| Increase projector dimension to 8192 (ablation) | MEC · ImageNet val | top-1 accuracy: 70.6 -> 71.2 | Improved |
Closest papers by meaning
- Learning Diverse and Discriminative Representations via the Principle of Maximal Coding Rate Reduction (NeurIPS 2020 · similarity 0.89)
- The Entropy Enigma: Success and Failure of Entropy Minimization (ICML 2024 · similarity 0.87)
- Matrix Information Theory for Self-Supervised Learning (ICML 2024 · similarity 0.87)
- Information Flow in Self-Supervised Learning (ICML 2024 · similarity 0.87)
- Cross-Entropy Is All You Need To Invert the Data Generating Process (ICLR 2025 · similarity 0.87)
- Learned Image Compression with Dictionary-based Entropy Model (CVPR 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.
