Research map / Architectures and efficiency / CNN design
Butterfly Transform: An Efficient FFT Based Neural Architecture Design
Keivan Alizadeh vahid, Anish Prabhu, Ali Farhadi, Mohammad Rastegari
Check 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 pointwise convolutions with BFT | ShuffleNetV2 · ImageNet validation | Top-1 accuracy: 50.86 -> 55.26 | Improved |
| Replace pointwise convolutions with BFT | MobileNetV1 · ImageNet validation | Top-1 accuracy: 41.50 -> 46.58 | Improved |
| Replace pointwise convolutions with BFT | MobileNetV3 · ImageNet validation | Top-1 accuracy: 49.8 -> 55.21 | Improved |
| Use BFT channel fusion instead of standard pointwise convolutions | ShuffleNetV2 at low FLOPs · ImageNet validation | Top-1 accuracy: 55.26 vs 50.2; 55.26 vs 49.3 | Improved |
| Use BFT channel fusion instead of alternative structured matrices | MobileNetV1 · ImageNet validation | Top-1 accuracy: 54.3 vs 49.65 | Improved |
| Add a first-to-last residual connection across Butterfly Layers (ablation) | MobileNetV1+BFT · CIFAR-10 | Accuracy: 79.2 -> 81.75 | Improved |
| Use first-to-last rather than every-other-layer residual connections (ablation) | MobileNetV1+BFT · CIFAR-10 | Accuracy: 81.12 -> 81.75 | Improved |
| Add nonlinear activations inside Butterfly Layers (ablation) | BFT · ImageNet classification | Accuracy | Hurt |
| Add batch normalization inside each Butterfly Layer (ablation) | BFT · ImageNet classification | FLOPs | Hurt |
| Apply weight decay to BFT weights (ablation) | BFT · ImageNet classification | Accuracy | Hurt |
| Use butterfly base k=4 rather than k=2, k=8, or k=n (ablation) | MobileNetV1+BFT · ImageNet, approximately 40M FLOPs | Accuracy | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- FBNet-96-0.35-1 — FBNet: Hardware-Aware Efficient ConvNet Design via Differentiable Neural Architecture Search (CVPR 2019) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Pixelated Butterfly: Simple and Efficient Sparse training for Neural Network Models (ICLR 2022 · similarity 0.85)
- ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices (CVPR 2018 · similarity 0.84)
- ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design (ECCV 2018 · similarity 0.84)
- LPViT: Low-Power Semi-structured Pruning for Vision Transformers (ECCV 2024 · similarity 0.84)
- RepViT: Revisiting Mobile CNN From ViT Perspective (CVPR 2024 · similarity 0.83)
- An Image is Worth More Than 16x16 Patches: Exploring Transformers on Individual Pixels (ICLR 2025 · similarity 0.83)
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.
