[Paper Review] ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices

Xiangyu Zhang, Xinyu Zhou+2 more|arXiv (Cornell University)|Jul 4, 2017

Advanced Neural Network Applications66 references882 citations

TL;DR

ShuffleNet introduces pointwise group convolutions with a channel shuffle operation to build extremely efficient CNNs for mobile devices, achieving higher accuracy than MobileNet at similar FLOPs and substantial real-world speedups on ARM hardware.

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.

Motivation & Objective

• Design a CNN architecture optimized for very low computational budgets (tens to hundreds of MFLOPs) for mobile devices.
• Reduce computation in 1x1 convolutions while maintaining accuracy.
• Enable effective information flow across channel groups via a channel shuffle mechanism.
• Demonstrate superior performance on ImageNet and MS COCO compared to contemporary lightweight models.

Proposed method

• Propose pointwise group convolution to reduce 1x1 convolution cost.
• Introduce channel shuffle to enable cross-group information flow across stacked group convolutions.
• Develop ShuffleNet units combining depthwise convolution on bottleneck channels with group-pointwise convolutions and channel shuffle.
• Assemble ShuffleNet into a three-stage architecture with staged downsampling and bottleneck channels set to 1/4 of stage output channels.
• Analyze computational complexity with FLOPs and compare across group counts (g) to study trade-offs between width and group sparsity.

Experimental results

Research questions

• RQ1 Does incorporating pointwise group convolutions with channel shuffle improve accuracy under strict FLOP budgets compared to existing lightweight architectures?
• RQ2 How does the number of groups (g) affect performance for tiny networks, and can channel shuffle mitigate potential bottlenecks in information flow?
• RQ3 Can ShuffleNet achieve higher accuracy than MobileNet at equivalent or lower computational costs on ImageNet and MS COCO?
• RQ4 What is the impact of ShuffleNet on actual mobile-device runtime compared to theoretical FLOP-based estimations?

Key findings

• ShuffleNet achieves superior accuracy vs. MobileNet at similar FLOPs on ImageNet, e.g., around 40 MFLOPs budget yields 7.8% lower top-1 error than MobileNet under comparable conditions.
• Using channel shuffle with pointwise group convolutions enables wider feature maps within the same complexity, improving accuracy especially for smaller networks.
• ShuffleNet 2x (g=3) attains 26.3% top-1 error at 524 MFLOPs, outperforming several compact architectures under the same budget.
• On ARM-based mobile hardware, ShuffleNet attains approximately 13x actual speedup over AlexNet, with comparable accuracy (theoretical speedups up to 18x).
• Channel shuffle consistently improves performance, particularly when a larger group count (g) is used, by enabling cross-group information flow.