Research map / Label-efficient and robust learning / Self-supervised and contrastive
Rethinking Supervised Pre-Training for Better Downstream Transferring
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 |
|---|---|---|---|
| Replace the supervised linear/MLP classifier with leave-one-out kNN pre-training | ResNet-50 pre-training · 9 fine-grained datasets, linear fine-tuning (mean) | accuracy: 75.88 -> 78.55 | Helped |
| Use LOOK rather than the strongest reported baseline | ResNet-50 pre-training · 9 fine-grained datasets, fully fine-tuning (mean) | accuracy: 88.72 -> 88.79 | Helped |
| Use LOOK rather than SimSiam pre-training | ResNet-50 · VOC 07+12 detection, two augmented views | APbb: 57.0 -> 56.3 | Hurt |
| Use two augmented views during LOOK pre-training (ablation) | LOOK · VOC 07+12 detection | APbb: 55.2 | Helped |
| Use strong rather than normal data augmentation (ablation) | LOOK · 9 fine-grained datasets, linear fine-tuning (mean) | accuracy: 77.60 -> 78.55 | Helped |
| Use strong rather than normal data augmentation (ablation) | LOOK · 9 fine-grained datasets, fully fine-tuning (mean) | accuracy: 88.03 -> 88.79 | Helped |
| Vary queue size from 65,536 to smaller sizes (ablation) | LOOK · 9 fine-grained datasets, linear fine-tuning (mean) | accuracy: 65,536: 78.55; 32,768: 78.23; 16,384: 77.72; 8,192: 77.71 | Hurt |
| Vary momentum across the tested values (ablation) | LOOK · 9 fine-grained datasets, linear fine-tuning (mean) | accuracy: 0.9999: 78.44; 0.999: 78.30; 0.99: 78.55; 0.9: 78.44 | No clear effect |
| Use fixed k=100 rather than the default decaying k (ablation) | LOOK · 9 fine-grained datasets, linear fine-tuning (mean) | accuracy: 75.19 | Hurt |
| Use fixed k=200 rather than the default decaying k (ablation) | LOOK · 9 fine-grained datasets, linear fine-tuning (mean) | accuracy: 78.42 | Hurt |
| Use fixed k=400 rather than the default decaying k (ablation) | LOOK · 9 fine-grained datasets, linear fine-tuning (mean) | accuracy: 77.93 | Hurt |
| Use fixed k=800 rather than the default decaying k (ablation) | LOOK · 9 fine-grained datasets, linear fine-tuning (mean) | accuracy: 77.51 | Hurt |
Closest papers by meaning
- Exploring the Limits of Large Scale Pre-training (ICLR 2022 · similarity 0.90)
- Learning What to Predict: Downstream-Guided Task Design for Continued Pretraining (NeurIPS 2026 · similarity 0.90)
- Understanding the Transferability of Representations via Task-Relatedness (NeurIPS 2024 · similarity 0.87)
- Knowledge Transfer from Vision Foundation Models for Efficient Training of Small Task-specific Models (ICML 2024 · similarity 0.87)
- Transferring Knowledge From Large Foundation Models to Small Downstream Models (ICML 2024 · similarity 0.87)
- Beyond Random Augmentations: Pretraining with Hard Views (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.