Identifying nonlinear dynamical systems via generative recurrent neural networks with applications to fMRI

Area of Science:

• Theoretical Neuroscience
• Computational Neuroscience
• Neuroimaging Analysis

Background:

• Cognitive and behavioral processes are implemented through neural system dynamics.
• Analyzing neurophysiological measurements aims to identify computational dynamics underlying task processing.

Purpose of the Study:

• To introduce a state space model (SSM) based on generative piecewise-linear recurrent neural networks (PLRNN) for assessing dynamics from neuroimaging data.
• To develop an interpretable and analyzable model for neural dynamics, adaptable to continuous-time systems.

Main Methods:

• A novel observation model for functional magnetic resonance imaging (fMRI) was coupled with a latent PLRNN.
• An efficient stepwise training procedure was implemented to prioritize capturing underlying dynamics over mere observation fitting.
• Kullback-Leibler divergence was used as an empirical measure to assess the approximation of underlying dynamics.

Main Results:

• The approach was validated on simulated and experimental fMRI data.
• Learned dynamics revealed task-related nonlinear structures missed by linear models.
• The model demonstrated the ability to capture 'true' underlying dynamics.

Conclusions:

• The developed PLRNN-based SSM offers an interpretable method for analyzing neural dynamics from fMRI.
• This approach can identify nonlinear structures in brain activity, advancing clinical assessment and neuroscientific research.
• It provides a novel step towards analyzing aberrant nonlinear dynamics using neuroimaging data.