Convolutional Neural Network Architecture Design Using an Improved Surrogate-Assisted Particle Swarm Optimization Algorithm

Optimizing Convolutional Neural Network (CNN) architectures significantly impacts image and natural language processing fields. Due to precise fitness evaluations, existing automatic design methods, especially those using evolutionary algorithms, are time-consuming. This paper introduces a new CNN design algorithm using particle swarm optimization (PSO), treating the design issue as a constrained optimization problem. PSO identifies optimal architectures within target parameters, utilizing a variable-length encoding strategy based on the ResNet BasicBlock (RB) to represent CNN architectures as particles of various sizes. A surrogate model replaces costly accuracy evaluations to expedite evolution. We use the CIFAR10 bench-mark data set to verify the proposed algorithm and compare it with some mainstream network structures and processes. Experimental results show that the classification accuracy of the proposed algorithm is equivalent to or even better than similar algorithms and consumes fewer computing resources.