AI Paper Atlas

Research map / Architectures and efficiency / CNN design

Efficient Training of Visual Transformers with Small Datasets

Yahui Liu, Enver Sangineto, Wei Bi, Nicu Sebe, Bruno Lepri, Marco De Nadai

NeurIPS 2021 · Architectures and efficiency › CNN design · approach: Compact CNN architectures · 84 citations (OpenAlex)

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 changedWhereMeasuredEffect
Add dense relative localization lossSwin-T · CIFAR-100, 100 epochs, from scratchTop-1 accuracy: 53.28 -> 66.23Improved
Add dense relative localization lossSwin-T · Quickdraw, 100 epochs, from scratchTop-1 accuracy: 24.08 -> 69.41Improved
Add dense relative localization lossSwin-T · ImageNet-100, 100 epochsTop-1 accuracy: 82.66 -> 83.95Improved
Add dense relative localization loss during fine-tuningSwin-T · Infograph, 100 epochs, fine-tuningTop-1 accuracy: 41.07 -> 46.10Improved
Add dense relative localization loss during pre-trainingSwin-T · ADE20K validation, ImageNet-1K pre-trainingmIoU: 43.87 -> 44.33Improved
Use signed target offsets instead of nonnegative offsets (ablation)Swin-T dense relative localization · ImageNet-100, 100 epochsTop-1 accuracy: 84.02 -> 83.14Hurt
Replace regression with cross-entropy classification (ablation)Swin-T dense relative localization · ImageNet-100, 100 epochsTop-1 accuracy: 84.02 -> 83.86Hurt
Use Gaussian-regularized localization loss (ablation)Swin-T dense relative localization · ImageNet-100, 100 epochsTop-1 accuracy: 84.02 -> 83.24Hurt
Apply localization loss at every transformer block (ablation)Swin-T dense relative localization · ImageNet-100, 100 epochsTop-1 accuracy: 82.76 -> 81.88Hurt
Use 32 rather than 64 embedding pairs per image (ablation)Swin-T dense relative localization · CIFAR-100, 100 epochsTop-1 accuracy: 66.23 -> 63.70Hurt
Use 128 rather than 64 embedding pairs per image (ablation)Swin-T dense relative localization · CIFAR-100, 100 epochsTop-1 accuracy: 66.23 -> 65.16Hurt
Use 256 rather than 64 embedding pairs per image (ablation)Swin-T dense relative localization · CIFAR-100, 100 epochsTop-1 accuracy: 66.23 -> 64.87Hurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy