Convolutional architecture search based on particle swarm algorithm for functional brain network classification

Abstract

The functional brain network (FBN) classification based on convolutional neural networks (CNN) is of great significance for discovery and diagnosis of brain diseases, and has attracted increasing attention. However, all the CNN architectures of current studies mainly depend on hand-crafted, which are labor-intensive and unreliable. To solve it, we propose a neural architecture search (NAS) method based on particle swarm optimization, to automatically design the CNN architecture for FBN classification. Specifically, this method includes three phases, namely the individual expression phase, the individual evaluation phase, and the individual update phase. In the first phase, we treat the neural architecture as the individual in particle swarm. The individual vector consists of six elements, and the value of each element represents the number of a special convolution operation. The six special convolution operations can effectively extract brain network multilevel topological features. In the second phase, we propose a novel surrogate-assisted predictor to evaluate the fitness of the individuals more efficiently. In the last phase, we apply the predicted fitness to acquire the historical optimum of each individual and the global optimum of the population, and use them to update all individuals in the particle swarm. The second and third phases are repeatedly performed until the end condition is met. Experiments on benchmark datasets demonstrate that the CNN architecture searched by our method achieves better classification performance than state-of-the-art hand-crafted CNN architectures.