期刊论文详细信息
IEEE Access 卷:9
Particle Swarm Optimization for Automatically Evolving Convolutional Neural Networks for Image Classification
Tom Lawrence1  Emma-Jane Phillips1  Chee Peng Lim2  Li Zhang3 
[1] Department of Computer and Information Sciences, Faculty of Engineering and Environment, Northumbria University, Newcastle upon Tyne, U.K;
[2] Institute for Intelligent Systems Research and Innovation, Deakin University, Waurn Ponds, VIC, Australia;
[3] National Subsea Centre, Robert Gordon University, Aberdeen, U.K;
关键词: Convolutional neural network;    encoding;    evolutionary computation;    image classification;    particle swarm optimization;   
DOI  :  10.1109/ACCESS.2021.3052489
来源: DOAJ
【 摘 要 】

Designing Convolutional Neural Networks from scratch is a time-consuming process that requires specialist expertise. While automated architecture generation algorithms have been proposed, the underlying search strategies generally are computationally expensive. The existing methods also do not explore the search space efficiently, and often lead to sub-optimal solutions. In this research, we propose a novel Particle Swarm Optimization (PSO)-based model for deep architecture generation to address the above challenges. Our proposed solution incorporates three new components. Firstly, a group-based encoding strategy is devised, which enforces the candidate networks to always follow the best practices. Specifically, it ensures that the number of groups can be adjusted in accordance with the input image size. By restricting the number of groups, we can adapt the frequency of the pooling operations toward the input image size. As such, it ascertains the position and maximum frequency of the pooling operations always result in a valid network architecture without the need for additional complex governing rules. Secondly, a new velocity updating mechanism is devised, which creates new network architectures by identifying the key network configuration differences. Thirdly, a new position updating mechanism using weighted velocity strengths is devised. Both the velocity and position updating mechanisms facilitate the proposed PSO-based model to search the intermediate positions of the particles' trajectories, allowing a better trade-off between diversification and intensification to be achieved. We employ eight well-known data sets, including Convex, Rectangles, MNIST and its variants, for model evaluation. The proposed PSO-based model achieves up to 7.58% improvement in accuracy and up to 63% reduction in computational cost, in comparison with those from the current state-of-the-art methods.

【 授权许可】

Unknown   

  文献评价指标  
  下载次数:0次 浏览次数:0次