Neural network acceleration methods via selective activation
Abstract The increase in neural network recognition accuracy is accompanied by a significant increase in the scales of networks and computations. To make deep learning frameworks widely used on mobile platforms, model acceleration has become extremely important in computer vision. In this study, a n...
| Main Authors: | , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Wiley
2023-04-01
|
| Series: | IET Computer Vision |
| Subjects: | |
| Online Access: | https://doi.org/10.1049/cvi2.12164 |
| _version_ | 1797846286528937984 |
|---|---|
| author | Siyu Wang WeiPeng Li Ruitao Lu Xiaogang Yang Jianxiang Xi Jiuan Gao |
| author_facet | Siyu Wang WeiPeng Li Ruitao Lu Xiaogang Yang Jianxiang Xi Jiuan Gao |
| author_sort | Siyu Wang |
| collection | DOAJ |
| description | Abstract The increase in neural network recognition accuracy is accompanied by a significant increase in the scales of networks and computations. To make deep learning frameworks widely used on mobile platforms, model acceleration has become extremely important in computer vision. In this study, a novel neural network acceleration method based on selective activation is proposed. First, as the algebraic basis for selective activation, mask general matrix multiplication is used to reduce matrix multiplication calculations. Second, to screen and remove activated neurons and reduce the number of calculations, we introduce an Activation Management Unit that includes two different strategies, Selective Activation with Primary Weights (SAPW) and Selective Activation with Primary Inputs (SAPI). SAPW greatly reduces the number of calculations of the fully connected layer and self‐attention and better guarantees detection accuracy. SAPI has the best performance on convolutional architectures, which can significantly reduce the amount of convolutional computation while maintaining the image classification accuracy. We present result of extensive experiments on computational and accuracy tradeoffs and show strong performance for CIFAR‐10 classification and Pascal VOC2012 detection. Compared with the dense method, the proposed selective activation method significantly reduces the number of neural network calculations with equal accuracy. |
| first_indexed | 2024-04-09T17:52:30Z |
| format | Article |
| id | doaj.art-5d8a358d3a844d3ea83feede4a5b0b8e |
| institution | Directory Open Access Journal |
| issn | 1751-9632 1751-9640 |
| language | English |
| last_indexed | 2024-04-09T17:52:30Z |
| publishDate | 2023-04-01 |
| publisher | Wiley |
| record_format | Article |
| series | IET Computer Vision |
| spelling | doaj.art-5d8a358d3a844d3ea83feede4a5b0b8e2023-04-15T11:16:51ZengWileyIET Computer Vision1751-96321751-96402023-04-0117329530810.1049/cvi2.12164Neural network acceleration methods via selective activationSiyu Wang0WeiPeng Li1Ruitao Lu2Xiaogang Yang3Jianxiang Xi4Jiuan Gao5Department of Automation PLA Rocket Force University of Engineering Xi'an ChinaCollege of Information and Communication National University of Defense Technology Wuhan ChinaDepartment of Automation PLA Rocket Force University of Engineering Xi'an ChinaDepartment of Automation PLA Rocket Force University of Engineering Xi'an ChinaDepartment of Automation PLA Rocket Force University of Engineering Xi'an ChinaDepartment of Automation PLA Rocket Force University of Engineering Xi'an ChinaAbstract The increase in neural network recognition accuracy is accompanied by a significant increase in the scales of networks and computations. To make deep learning frameworks widely used on mobile platforms, model acceleration has become extremely important in computer vision. In this study, a novel neural network acceleration method based on selective activation is proposed. First, as the algebraic basis for selective activation, mask general matrix multiplication is used to reduce matrix multiplication calculations. Second, to screen and remove activated neurons and reduce the number of calculations, we introduce an Activation Management Unit that includes two different strategies, Selective Activation with Primary Weights (SAPW) and Selective Activation with Primary Inputs (SAPI). SAPW greatly reduces the number of calculations of the fully connected layer and self‐attention and better guarantees detection accuracy. SAPI has the best performance on convolutional architectures, which can significantly reduce the amount of convolutional computation while maintaining the image classification accuracy. We present result of extensive experiments on computational and accuracy tradeoffs and show strong performance for CIFAR‐10 classification and Pascal VOC2012 detection. Compared with the dense method, the proposed selective activation method significantly reduces the number of neural network calculations with equal accuracy.https://doi.org/10.1049/cvi2.12164computer visionconvolutional neural netsimage processingneural net architecture |
| spellingShingle | Siyu Wang WeiPeng Li Ruitao Lu Xiaogang Yang Jianxiang Xi Jiuan Gao Neural network acceleration methods via selective activation IET Computer Vision computer vision convolutional neural nets image processing neural net architecture |
| title | Neural network acceleration methods via selective activation |
| title_full | Neural network acceleration methods via selective activation |
| title_fullStr | Neural network acceleration methods via selective activation |
| title_full_unstemmed | Neural network acceleration methods via selective activation |
| title_short | Neural network acceleration methods via selective activation |
| title_sort | neural network acceleration methods via selective activation |
| topic | computer vision convolutional neural nets image processing neural net architecture |
| url | https://doi.org/10.1049/cvi2.12164 |
| work_keys_str_mv | AT siyuwang neuralnetworkaccelerationmethodsviaselectiveactivation AT weipengli neuralnetworkaccelerationmethodsviaselectiveactivation AT ruitaolu neuralnetworkaccelerationmethodsviaselectiveactivation AT xiaogangyang neuralnetworkaccelerationmethodsviaselectiveactivation AT jianxiangxi neuralnetworkaccelerationmethodsviaselectiveactivation AT jiuangao neuralnetworkaccelerationmethodsviaselectiveactivation |