Prioritized learning of cross-population neural dynamics.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Systems Neuroscience

Background:

• Advancements in multi-region recording technology allow for studying interactions between distinct brain regions.
• A significant computational challenge is distinguishing cross-regional neural dynamics from within-region dynamics, which can mask or confound the analysis.

Purpose of the Study:

• To introduce a novel computational framework, cross-population prioritized linear dynamical modeling (CroP-LDM), designed to address the challenge of modeling cross-regional neural dynamics.
• To enable accurate learning and inference of latent states representing cross-population dynamics, unconfounded by within-population activity.

Main Methods:

• CroP-LDM employs a prioritized learning approach to model cross-population dynamics using latent states.
• The method allows for both causal (using past data) and non-causal temporal inference of these latent states.
• Validation involved comparisons with existing linear dynamical modeling (LDM) methods and application to multi-regional neural recordings.

Main Results:

• The prioritized learning objective in CroP-LDM was identified as crucial for accurately learning cross-population dynamics.
• CroP-LDM demonstrated superior performance in learning cross-population dynamics compared to static and dynamic methods, even with low-dimensional data, using motor and premotor cortical recordings.
• The approach successfully quantified dominant interaction pathways across brain regions in an interpretable manner.

Conclusions:

• CroP-LDM provides a robust framework for analyzing neural dynamics across multiple brain regions.
• The method effectively overcomes the confounding effects of within-region activity, offering a significant advancement in computational neuroscience.
• CroP-LDM facilitates a deeper understanding of inter-regional brain communication during complex tasks.