Evolutional Neural Architecture Search for Optimization of Spatiotemporal Brain Network Decomposition

Area of Science:

• Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Analyzing human brain activity from 4D fMRI data presents challenges in designing effective deep neural networks (DNNs).
• Previous methods like DBN, CNN, and DSRAE rely on manual architecture and hyperparameter selection, limiting optimality.
• Manual DNN design for spatiotemporal pattern extraction from fMRI is complex and suboptimal.

Purpose of the Study:

• To develop an automated framework for optimizing deep neural network architectures for fMRI data analysis.
• To introduce evolutionary algorithms for neural architecture search (NAS) in the context of brain activity analysis.
• To enhance the extraction of meaningful spatiotemporal patterns from 4D fMRI data.

Main Methods:

• Employed evolutionary algorithms (EA) to optimize the deep neural architecture of a Deep Sparse Recurrent Auto-Encoder (DSRAE).
• Developed an automated neural architecture search (NAS) method, termed eNAS-DSRAE, by minimizing the expected loss of initialized models.
• Validated the eNAS-DSRAE framework on the Human Connectome Project (HCP) 900 datasets.

Main Results:

• The eNAS-DSRAE framework successfully identified spatiotemporal features from 4D fMRI data.
• Optimized eNAS-DSRAE demonstrated superior performance compared to manually designed neural network models.
• The framework effectively extracts connectome-scale meaningful spatiotemporal brain networks.

Conclusions:

• eNAS-DSRAE provides an effective, automated approach for optimizing deep neural networks for fMRI data.
• This work represents an early application of NAS for extracting brain networks from fMRI data.
• The proposed framework is effective for optimizing recurrent neural network (RNN)-based models for neuroimaging analysis.