A Novel GPU-Based Efficient Approach for Convolutional Neural Networks with Small Filters

详细信息    查看全文

• 作者：Wenbin Jiang ; Yiming Chen ; Hai Jin ; Ran Zheng…
• 关键词：Deep neural networks ; Convolutional neural network ; Parallel computation ; GPU
• 刊名：Journal of Signal Processing Systems
• 出版年：2017
• 出版时间：March 2017
• 年：2017
• 卷：86
• 期：2-3
• 页码：313-325
• 全文大小：
• 刊物类别：Engineering
• 刊物主题：Signal,Image and Speech Processing; Circuits and Systems; Electrical Engineering; Image Processing and Computer Vision; Pattern Recognition; Computer Imaging, Vision, Pattern Recognition and Graphics;
• 出版者：Springer US
• ISSN：1939-8115
• 卷排序：86

文摘

In recent years, convolutional neural networks (CNNs) as important parts of deep neural networks (DNNs) have achieved great successes in the field of computer vision. However, Convolution always takes much computation time in the DNNs. In order to improve the efficiency of CNNs, many solutions focusing on training algorithms and parallelism strategies have been proposed. In this paper, different from traditional GPU-based algorithms, a novel algorithm based on look-up table is proposed to speed up the CNNs with small filters by applying GPU. By transforming complex matrix multiplications operations in the convolution computation to some table-based simple summation operations, the overhead of convolution computation can be considerably reduced. The process of creating a table and looking up values in the table is very appropriate for parallelization on a GPU. The experimental results show that the proposed approach can improve the speed of convolution computation by 20–30 %, compared with existing state-of-the-art works with less accuracy loss.