AI Paper Atlas

Research map / Architectures and efficiency / CNN design

Deformable Butterfly: A Highly Structured and Sparse Linear Transform

Rui Lin, Jie Ran, King Hung Chiu, Graziano Chesi, Ngai Wong

NeurIPS 2021 · Architectures and efficiency › CNN design · approach: CNN filter pruning · 5 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
Replace FC/CONV layers CONV2+FC1+FC2 with DeBut chainsmodified LeNet · MNISTaccuracy (%): 99.29 -> 98.64Hurt
Add ALS initialization to DeBut chains (ablation)DeBut CONV2+FC1+FC2 · MNISTaccuracy (%): 98.64 -> 97.27Hurt
Add ALS initialization to DeBut chain on CONV2 (ablation)DeBut CONV2 · MNISTaccuracy (%): 99.02 -> 98.86Hurt
Replace Butterfly with DeBut on three layersLeNet · MNISTaccuracy (%): 98.02 -> 98.64Improved
Replace Adaptive Fastfood with DeBut on FC1LeNet · MNISTaccuracy (%): 98.89 -> 98.89No clear effect
Replace CONV8-13 of VGG-16-BN with DeBut chainsVGG-16-BN · CIFAR-10accuracy (%): 93.96 -> 93.72Hurt
Replace last CONV13 only with DeBut chainVGG-16-BN · CIFAR-10accuracy (%): 93.96 -> 93.91No clear effect
Replace Adaptive Fastfood with DeButVGG-16-BN CONV8-13 · CIFAR-10training time (s/epoch): 2100 -> 50Improved
Replace Butterfly with DeButVGG-16-BN CONV8-13 · CIFAR-10accuracy (%): 93.34 -> 93.72Improved
Replace Tucker-2 low-rank decomposition with DeButVGG-16-BN CONV8-13 · CIFAR-10accuracy (%): 93.36 -> 93.71Improved
Use monotonic instead of bulging chains (ablation)DeBut ResNet-50 · ImageNettop-1 accuracy (%): 74.52 -> 74.34No clear effect
Use monotonic instead of bulging chains (ablation)DeBut ResNet-50 · ImageNettop-5 accuracy (%): 92.18 -> 92.31No 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