Research map / Label-efficient and robust learning / Recognition and long-tail
Train a One-Million-Way Instance Classifier for Unsupervised Visual Representation Learning
Read the paperCheck 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 |
|---|---|---|---|
| Use full instance classification rather than sampling 2^16 classes (ablation) | Instance classifier · ImageNet-1K linear evaluation | Top-1: 65.5 -> 67.3 | Helped |
| Add raw-feature classifier initialization using a fixed random network with running BN (ablation) | Instance classifier · ImageNet-1K, 10 epochs | Top-1: 12.4 -> 27.4 (+15.0) | Helped |
| Add raw-feature classifier initialization using a fixed random network with running BN (ablation) | Instance classifier · ImageNet-1K, 25 epochs | Top-1: 40.8 -> 46.3 (+5.5) | Helped |
| Add raw-feature classifier initialization using a fixed random network with running BN (ablation) | Instance classifier · ImageNet-1K, 50 epochs | Top-1: 56.1 -> 58.0 (+1.9) | Helped |
| Add raw-feature classifier initialization using a fixed random network with running BN (ablation) | Instance classifier · ImageNet-1K, 100 epochs | Top-1: 62.9 -> 64.1 (+1.2) | Helped |
| Add raw-feature classifier initialization using a fixed random network with running BN (ablation) | Instance classifier · ImageNet-1K, 200 epochs | Top-1: 67.3 -> 67.6 (+0.3) | Helped |
| Add raw-feature classifier initialization using a fixed random network with running BN (ablation) | Instance classifier · ImageNet-1K, 400 epochs | Top-1: 69.3 -> 69.7 (+0.4) | Helped |
| Add hardest-class label smoothing with K=100 and alpha=0.2 (ablation) | Instance classifier with contrastive prior · ImageNet-1K linear evaluation | Top-1: 67.6 -> 68.2 | Helped |
| Increase smoothing factor to alpha=0.3 for K=100 (ablation) | Instance classifier · ImageNet-1K linear evaluation | Top-1: 68.2 -> 67.5 | Hurt |
| Add a two-layer MLP head (ablation) | Instance classification with contrastive prior and label smoothing · ImageNet-1K linear evaluation | Top-1: 58.6 -> 68.2 | Helped |
| Add contrastive-prior initialization (ablation) | Instance classifier with MLP head · ImageNet-1K linear evaluation | Top-1: 67.3 -> 67.6 | Helped |
| Add hardest-class label smoothing (ablation) | Instance classifier with MLP head and contrastive prior · ImageNet-1K linear evaluation | Top-1: 67.6 -> 68.2 | Helped |
Closest papers by meaning
- Unsupervised Feature Learning via Non-Parametric Instance Discrimination (CVPR 2018 · similarity 0.91)
- Parametric Instance Classification for Unsupervised Visual Feature learning (NeurIPS 2020 · similarity 0.91)
- Revisiting Self-Supervised Visual Representation Learning (CVPR 2019 · similarity 0.89)
- Unsupervised Representation Learning by Predicting Image Rotations (ICLR 2018 · similarity 0.88)
- Beyond Supervised vs. Unsupervised: Representative Benchmarking and Analysis of Image Representation Learning (CVPR 2022 · similarity 0.88)
- Self-Organizing Visual Prototypes for Non-Parametric Representation Learning (ICML 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.