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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| Replace FC/CONV layers CONV2+FC1+FC2 with DeBut chains | modified LeNet · MNIST | accuracy (%): 99.29 -> 98.64 | Hurt |
| Add ALS initialization to DeBut chains (ablation) | DeBut CONV2+FC1+FC2 · MNIST | accuracy (%): 98.64 -> 97.27 | Hurt |
| Add ALS initialization to DeBut chain on CONV2 (ablation) | DeBut CONV2 · MNIST | accuracy (%): 99.02 -> 98.86 | Hurt |
| Replace Butterfly with DeBut on three layers | LeNet · MNIST | accuracy (%): 98.02 -> 98.64 | Improved |
| Replace Adaptive Fastfood with DeBut on FC1 | LeNet · MNIST | accuracy (%): 98.89 -> 98.89 | No clear effect |
| Replace CONV8-13 of VGG-16-BN with DeBut chains | VGG-16-BN · CIFAR-10 | accuracy (%): 93.96 -> 93.72 | Hurt |
| Replace last CONV13 only with DeBut chain | VGG-16-BN · CIFAR-10 | accuracy (%): 93.96 -> 93.91 | No clear effect |
| Replace Adaptive Fastfood with DeBut | VGG-16-BN CONV8-13 · CIFAR-10 | training time (s/epoch): 2100 -> 50 | Improved |
| Replace Butterfly with DeBut | VGG-16-BN CONV8-13 · CIFAR-10 | accuracy (%): 93.34 -> 93.72 | Improved |
| Replace Tucker-2 low-rank decomposition with DeBut | VGG-16-BN CONV8-13 · CIFAR-10 | accuracy (%): 93.36 -> 93.71 | Improved |
| Use monotonic instead of bulging chains (ablation) | DeBut ResNet-50 · ImageNet | top-1 accuracy (%): 74.52 -> 74.34 | No clear effect |
| Use monotonic instead of bulging chains (ablation) | DeBut ResNet-50 · ImageNet | top-5 accuracy (%): 92.18 -> 92.31 | No clear effect |
Closest papers by meaning
- Pixelated Butterfly: Simple and Efficient Sparse training for Neural Network Models (ICLR 2022 · similarity 0.89)
- Learning Fast Algorithms for Linear Transforms Using Butterfly Factorizations (ICML 2019 · similarity 0.87)
- ButterflyFlow: Building Invertible Layers with Butterfly Matrices (ICML 2022 · similarity 0.85)
- Searching for Efficient Linear Layers over a Continuous Space of Structured Matrices (NeurIPS 2024 · similarity 0.84)
- Dense Structural Compression of Transformers via Gauge-Correct Channel Removal (NeurIPS 2026 · similarity 0.83)
- FrameQuant: Flexible Low-Bit Quantization for Transformers (ICML 2024 · similarity 0.82)
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.
