CBP: Learning Shared Cognitive Basis Space and Connectivity Patterns for Cross-Task Brain Dynamics Modeling

Weibin Li, Wendu Li, Xihua Yin, Zongxiang Tan, Yushan You, Xinke Shen, Quanying Liu

Keywords: Cross-Task Brain Dynamics, Cognitive mechanisms, Neural dynamics

Abstract: Understanding the brain dynamics underlying human cognition requires models that jointly achieve modeling capability and interpretability across tasks. Existing approaches either rely on deep learning models, which learn strong but implicitly encoded representations that are difficult to interpret, or explicit state-space models, which are interpretable but limited to single-task settings. To bridge this gap, we present CBP, an interpretable, non–deep-learning framework for cross-cognitive-task brain modeling. By explicitly leveraging shared information across tasks, CBP accurately recovers latent components while maintaining state-of-the-art performance. Importantly, CBP uncovers a stable set of \underline{\textbf{C}}ognitive \underline{\textbf{B}}ases and connectivity \underline{\textbf{P}}atterns in the human brain. The reliability of these discoveries is supported by extensive quantitative evaluations and a battery of perturbation analyses, demonstrating their robustness. Moreover, leveraging these shared cognitive components allows CBP to generalize effectively to both the HCP and multi-stage learning datasets, where it not only achieves accurate prediction of task states and learning outcomes but also reveals the key cognitive connectivity patterns underlying these behaviors. Together, these results establish CBP as an interpretable framework for large-scale cognitive dynamics, offering mechanistic insight into cross-task cognition.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 6762