ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices

TL;DR

ShuffleNet is a highly efficient CNN architecture optimized for mobile devices, employing novel operations to reduce computation while maintaining high accuracy, and demonstrating significant speedups over traditional models.

Contribution

Introduces ShuffleNet, a new CNN architecture with pointwise group convolution and channel shuffle, achieving high efficiency and accuracy for mobile applications.

Findings

Outperforms MobileNet with 7.8% lower top-1 error on ImageNet at 40 MFLOPs

Achieves ~13x speedup over AlexNet on ARM devices

Maintains comparable accuracy with significantly reduced computation

Abstract

We introduce an extremely computation-efficient CNN architecture named ShuffleNet, which is designed specially for mobile devices with very limited computing power (e.g., 10-150 MFLOPs). The new architecture utilizes two new operations, pointwise group convolution and channel shuffle, to greatly reduce computation cost while maintaining accuracy. Experiments on ImageNet classification and MS COCO object detection demonstrate the superior performance of ShuffleNet over other structures, e.g. lower top-1 error (absolute 7.8%) than recent MobileNet on ImageNet classification task, under the computation budget of 40 MFLOPs. On an ARM-based mobile device, ShuffleNet achieves ~13x actual speedup over AlexNet while maintaining comparable accuracy.