Adaptive particle swarm architecture search based on multi-level convolutions for functional brain network classification

Recently, the functional brain network (FBN) classification methods based on deep neural networks (DNNs) have around a lot of scientific interest. However, these DNN architectures are manually designed by human experts through trial-and-error testing, which not only requires rich parameter tuning experience and large labor costs, but also a fixed manual architecture cannot consistently guarantee good performance across different data distributions and scenarios. To solve this problem, we propose an adaptive particle swarm architecture search method based on multi-level convolutions, which can automatically design suitable DNN architectures for various FBN classification tasks. Specifically, to effectively extract multi-level features at FBN, we construct three multi-level convolution units to form candidate architectures. These units can extract edge-level, node-level, and graph-level features respectively. The parameters of these units will be searched using the particle swarm-based NAS framework. Additionally, to alleviate the difficulty of searching in a vast search space, we propose a novel adaptive updating strategy. This strategy adaptively locks specific elements of the particle vector based on historical information and the search epochs, which can effectively search within a subset of the vast search space. We conduct systematic experiments on ABIDE I, ABIDE II, and ADHD-200 datasets with different atlases. The experimental results demonstrate that our method achieves competitive accuracies of 74.71%, 73.03%, and 74.39% on the CC200 atlas, and 71.42%, 73.91%, and 69.96% on the AAL atlas respectively.