Adaptive latent state modeling of brain network dynamics with real-time learning rate optimization.

Area of Science:

• Computational neuroscience
• Neuroimaging analysis
• Dynamic systems modeling

Background:

• Dynamic latent state models are crucial for analyzing brain network activity.
• Existing models often assume stationary dynamics, failing to capture non-stationarities like learning or instability.
• Modeling non-stationary brain dynamics requires adaptive parameter updates and optimized learning rates.

Purpose of the Study:

• To develop methods for adaptively updating parameters in latent state models despite the latent nature of states.
• To create algorithms for optimizing the learning rate in adaptive models for improved accuracy.
• To introduce a unified algorithm, RO-adaptive LSSM, for modeling non-stationary brain network dynamics.

Main Methods:

• Developed a computation- and memory-efficient adaptive Linear State-Space Modeling (LSSM) fitting algorithm for real-time recursive parameter updates.
• Designed a real-time learning rate optimization algorithm to work in parallel with the adaptive LSSM fitting.
• Validated the RO-adaptive LSSM algorithm using comprehensive simulations of diverse non-stationary brain network dynamics.

Main Results:

• The adaptive LSSM fitting algorithm accurately tracked simulated non-stationary brain network dynamics.
• The learning rate was found to significantly impact LSSM fitting accuracy.
• The RO-adaptive LSSM algorithm demonstrated rapid convergence to optimal learning rates and accurate tracking of non-stationarities.

Conclusions:

• The RO-adaptive LSSM algorithm effectively models time-varying neural dynamics.
• These algorithms enhance the study of brain functions and improve neurotechnologies like brain-machine interfaces.
• This work provides a robust framework for analyzing dynamic and non-stationary brain activity.