PDCNN-MRW: a parallel Winograd convolutional neural network algorithm base on MapReduce
Parallel deep convolutional neural network (DCNN) algorithms have been widely used in the field of big data, but there are still some problems: excessive computation of redundant features, insufficient performance of convolution operation, and poor merging ability of parameter parallelization. Based...
Saved in:
Published in | International journal of machine learning and cybernetics Vol. 15; no. 5; pp. 1949 - 1966 |
---|---|
Main Authors | , , |
Format | Journal Article |
Language | English |
Published |
Berlin/Heidelberg
Springer Berlin Heidelberg
01.05.2024
Springer Nature B.V |
Subjects | |
Online Access | Get full text |
Cover
Loading…
Summary: | Parallel deep convolutional neural network (DCNN) algorithms have been widely used in the field of big data, but there are still some problems: excessive computation of redundant features, insufficient performance of convolution operation, and poor merging ability of parameter parallelization. Based on the above problems, a parallel DCNN algorithm based on MapReduce and Winograd convolution (PDCNN-MRW) is proposed in this paper, which contains three parts. First, a feature selection method based on cosine similarity and normalized mutual information (FS-CSNMI) is proposed, which reduces redundant feature computation between channels and avoids excessive redundant feature computation. Next, a parallel Winograd convolution method base on MapReduce (PWC-MR) is presented to address the insufficient convolution performance by reducing the number of multiplications. Finally, a load balancing method based on multiway tree and task migration (LB-MTTM) is developed, which improves the capability of parameter merging by balancing the load between nodes and reducing the response time of the cluster. We compared the PDCNN-MRW algorithm with other algorithms on four datasets, including ISIC 2019, BloodCellImages, PatchCamelyon, and ImageNet 1K. Experiment shows that the proposed algorithm has lower training costs and higher efficiency than other parallel DCNN algorithms. |
---|---|
ISSN: | 1868-8071 1868-808X |
DOI: | 10.1007/s13042-023-02007-0 |