Evolving convolutional neural networks by symbiotic organisms search algorithm for image classification

Abstract Convolutional neural networks (CNNs) are important neural networks in the deep learning field. In general, CNN design requires manual iterations and adjustments based on the characteristics of the data and a priori knowledge of the application area, and such processes are time-consuming, labor-intensive, and weaken the generalizability of the CNN architecture. Due to the complexity of the CNN architecture, its automatic construction is difficult. Automating the design of a CNN architecture using evolutionary algorithms is viable, but the field is still in its infancy. We propose an architectural search algorithm for CNNs that employs a symbiotic organisms search (SOS) algorithm called sosCNN. First, a SOS algorithm with strong global optimization abilities was introduced for CNN architecture search, and this algorithm facilitates the automatic building of good CNN architectures. Then, a novel integrated coding update method is proposed that reduces the loss of convolutional layers, which results in a searched architecture with stronger feature extraction capability. Finally, three new non-numeric computational strategies, namely, binary segmentation, slack gain, and dissimilar mutation, were combined with the SOS algorithm. We tested the sosCNN algorithm against 24 algorithms, including state-of-the-art algorithms, on nine widely used image classification datasets. The experimental results show that the average classification error of sosCNN is reduced by 0.04% to 7.37% on the nine benchmark test sets compared to the state of art algorithms, which is a very promising result.