Research map / Architectures and efficiency / CNN design
Do Better ImageNet Models Transfer Better?
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 |
|---|---|---|---|
| use a higher ImageNet-accuracy architecture | logistic regression on fixed penultimate-layer features · 12 image classification datasets | transfer accuracy: r = 0.99 | Helped |
| use a higher ImageNet-accuracy architecture | fine-tuned CNNs · 12 image classification datasets | transfer accuracy: r = 0.96 | Helped |
| use a higher ImageNet-accuracy architecture | CNNs trained from random initialization · 12 image classification datasets | transfer accuracy: r = 0.55 | Helped |
| add batch-normalization scale parameters (ablation) | Inception models used as fixed feature extractors · image classification transfer datasets | transfer accuracy | Helped |
| disable label smoothing (ablation) | ImageNet-pretrained Inception models used as fixed feature extractors · image classification transfer datasets | transfer accuracy | Helped |
| disable dropout (ablation) | ImageNet-pretrained Inception models used as fixed feature extractors · image classification transfer datasets | transfer accuracy | Helped |
| remove the auxiliary classifier head (ablation) | ImageNet-pretrained Inception models used as fixed feature extractors · image classification transfer datasets | transfer accuracy | Helped |
| use regularization during ImageNet pretraining (ablation) | logistic regression on public-checkpoint features · 12 image classification datasets | Pearson r between ImageNet and transfer accuracy: 0.99 -> 0.37 | Hurt |
| fine-tune ImageNet-pretrained weights rather than train from scratch (ablation) | Inception v4 · Stanford Cars | top-1 accuracy: 92.7% -> 93.3% | Helped |
| fine-tune ImageNet-pretrained weights rather than train from scratch (ablation) | Inception v4 · FGVC Aircraft | mean per-class accuracy: 88.8% -> 89.0% | Helped |
| initialize from ImageNet weights rather than random initialization (ablation) | Inception v4 · fine-grained classification tasks | steps to reach 90% of maximum odds of correct classification: 1151 steps -> 19531 steps | Helped |
| add the auxiliary head during pretraining and fine-tuning (ablation) | Inception v4 · Food-101, FGVC Aircraft, VOC2007, Oxford Pets, CIFAR-100, DTD, Oxford 102 Flowers | fine-tuning accuracy | Mixed |
Closest papers by meaning
- Does progress on ImageNet transfer to real-world datasets? (NeurIPS 2023 · similarity 0.92)
- Do Adversarially Robust ImageNet Models Transfer Better? (NeurIPS 2020 · similarity 0.90)
- Why Do Better Loss Functions Lead to Less Transferable Features? (NeurIPS 2021 · similarity 0.89)
- Bag of Tricks for Image Classification with Convolutional Neural Networks (CVPR 2019 · similarity 0.89)
- Does Robustness on ImageNet Transfer to Downstream Tasks? (CVPR 2022 · similarity 0.88)
- ConvNet vs Transformer, Supervised vs CLIP: Beyond ImageNet Accuracy (ICML 2024 · 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.