AI Paper Atlas

Research map / Architectures and efficiency / Pruning and efficient tuning

Pixelated Butterfly: Simple and Efficient Sparse training for Neural Network Models

Beidi Chen, Tri Dao, Kaizhao Liang, Jiaming Yang, Zhao Song, Atri Rudra, Christopher Re

ICLR 2022 · Architectures and efficiency › Pruning and efficient tuning · approach: Sparse training from scratch · 7 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
use Pixelated Butterfly sparse trainingMLP-Mixer and ViT · ImageNet, CIFAR-10/100training speedup: up to 2.3×Improved
use Pixelated ButterflyGPT-2 small · WikiText-103training speedup: 2.1×Improved
use Pixelated ButterflyGPT-2 medium · WikiText-103training speedup: up to 2.5×Improved
use Pixelated ButterflyTransformer · Long Range Arenatraining speedup: 5.2×Improved
replace existing hand-crafted sparsity patterns with butterfly sparsity (ablation)Pixelated Butterfly · ImageNetaccuracy: up to 2%Improved
align sparsity to hardware blocks (ablation)Pixelated Butterfly · Sparse trainingtraining speedup: up to 5×Improved
allocate compute budget across all layers according to their compute share (ablation)ViT-small · ViT-smalltraining speedup: 2×Improved
allocate about one-quarter of the budget to low-rank and the remainder to flat block butterfly (ablation)Pixelated Butterfly Mixer-S · CIFAR-10accuracyImproved
use flat block butterfly instead of regular butterfly (ablation)Pixelated Butterfly · GPU multiplicationmultiplication speed: up to 3×Improved
sparsify both attention and MLP rather than only one component (ablation)ViT-small · ViT-smallend-to-end training speedupImproved
use the smallest hardware-supported block size (ablation)Pixelated Butterfly · Sparse matrix multiplicationtraining speedImproved
use Pixelated Butterfly instead of dense trainingMixer-S/32 · CIFAR-10accuracyImproved

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