Artificial Bee Colony Optimization based Optimal Convolutional Neural Network Architecture Design

Convolutional neural networks (CNNs) are broadly used to solve various computer vision tasks. However, designing optimal CNN architecture for solving a particular task is a trial- and-error process and thus requires domain-specific expertise. Recently, several meta-heuristic strategies have been developed to solve the design challenges of CNN models. Here, we propose an approach to designing optimal CNN architecture using the artificial bee colony (ABC) algorithm and named as ABC-CNN. A variable-length encoding scheme and refinement strategy are proposed to evolve CNN architectures with the ABC algorithm. The proposed approach is evaluated on three well-known image classification datasets. Experimental results demonstrate the effectiveness of the proposed ABC-CNN algorithm compared with 11 state-of-the-art models, including manual and meta-heuristic based CNN models.