Neural Architecture Search for Optimization of Spatial-Temporal Brain Network Decomposition

Using neural networks to explore spatial patterns and temporal dynamics of human brain activities has been an important yet challenging problem because it is hard to manually design the most optimal neural networks. There have been several promising deep learning methods that can decompose neuroscientifically meaningful spatial-temporal patterns from 4D fMRI data, e.g., the deep sparse recurrent auto-encoder (DSRAE). However, those previous studies still depend on hand-crafted neural network structures and hyperparameters, which are not optimal in various senses. In this paper, we employ evolutionary algorithms to optimize such DSRAE neural networks by minimizing the expected loss of the generated architectures on data reconstruction via the neural architecture search (NAS) framework, named NAS-DSRAE. The optimized NAS-DSRAE is evaluated by the publicly available human connectome project (HCP) fMRI datasets and our promising results showed that NAS-DSRAE has sufficient generalizability to model the spatial-temporal features and is better than the hand-crafted model. To our best knowledge, the proposed NAS-DSRAE is among the earliest NAS models that can extract connectome-scale meaningful spatial-temporal brain networks from 4D fMRI data.