Modeling functional brain networks is essential for revealing the functional mechanisms of the human brain. Deep Neural Network (DNN) models have been widely employed for extracting multi-scale spatiotemporal features from functional magnetic resonance imaging (fMRI) data. However, current approaches face two fundamental challenges. Firstly, existing deep neural network-based approaches exhibit significant limitations in learning cross-task common representations when dealing with the variable sequence length characteristics inherent in task-fMRI multi-task data. Secondly, existing approaches often neglect the dynamic variability of neural activity across different time points in fMRI data. To overcome these challenges, a novel framework based on \(\:{\mathbf{L}}_{2}\) -Normalized Attention Fully Convolutional Recurrent Autoencoder ( \(\:{\mathbf{L}}_{2}\) -FCRAAE) is proposed for modeling hierarchical functional brain networks (FBNs). Specifically, the \(\:{\mathbf{L}}_{2}\) -FCRAAE is trained in an unsupervised manner, where the autoencoder architecture guides the attention modules to focus on task-activated regions. The architecture incorporates two synergistic design principles: First, its fully convolutional recurrent structure inherently adapts to variable-length tfMRI time series while effectively capturing long-range temporal dynamics and recognizing brain state transitions. Second, the integrated \(\:{\mathbf{L}}_{2}\) -normalized temporal-channel attention module weights task-relevant neural activation patterns, substantially enhancing representational capacity. Comprehensive experiments demonstrate that the proposed \(\:{\mathbf{L}}_{2}\) -FCRAAE exhibits superior capability and generalizability in characterizing spatial and temporal patterns of FBNs in a hierarchical manner. Overall, this study presents a novel approach for understanding the hierarchical organization of functional brain architecture. The code for this paper is available at: https://github.com/beiweizai111/FCAAE.
