AI Paper Atlas

Research map / Architectures and efficiency / CNN design

Approximating CNNs with Bag-of-local-Features models works surprisingly well on ImageNet

Wieland Brendel, Matthias Bethge

ICLR 2019 · Architectures and efficiency › CNN design · approach: Efficient CNN scaling and dynamic inference · 141 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
Limit classification to local features within 33 × 33 pixel patches and average patch evidence linearlyBagNet-33 · ImageNettop-5 accuracy: 87.6%Improved
Limit classification to local features within 17 × 17 pixel patches and average patch evidence linearlyBagNet-17 · ImageNettop-5 accuracy: 80.5%Improved
Use BagNet inference instead of ResNet-50 inferenceBagNet-9, BagNet-17, BagNet-33 · ImageNet, 3 × 224 × 224 images, batch size 64inference throughput: around 155 images/s vs 570 images/sHurt
Texturise images while preserving local feature statisticsVGG-16 · ImageNet texturised imagesaccuracy: 90.1% on clean versus 79.4% on texturised imageHurt
Use BagNet-17 heatmaps to select patches for maskingVGG-16 · ImageNet, masking 100 highest-attribution 8 × 8 patchesaverage probability of leading class: 0.151Improved
Use BagNet-17 heatmaps to select patches for maskingResNet-50 · ImageNet, masking 100 highest-attribution 8 × 8 patchesaverage probability of leading class: 0.263Improved
Use BagNet-17 heatmaps to select patches for maskingResNet-152 · ImageNet, masking 100 highest-attribution 8 × 8 patchesaverage probability of leading class: 0.371Improved
Use BagNet-17 heatmaps to select patches for maskingDenseNet-169 · ImageNet, masking 100 highest-attribution 8 × 8 patchesaverage probability of leading class: 0.326Improved
Replace patch logits below a threshold with the threshold value (ablation)BagNet-33 · ImageNettop-5 accuracyNo clear effect
Binarize patch heatmaps by thresholding their values (ablation)BagNet-33 · ImageNettop-5 accuracy: within 3-4% of the vanilla BagNet performanceNo clear effect

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