Meta-Dynamical State Space Models for Integrative Neural Data Analysis.

Area of Science:

• Computational neuroscience
• Machine learning
• Systems neuroscience

Background:

• Shared structure learning enhances neural and machine learning adaptability.
• Existing methods struggle with statistical variations across neural recordings.
• Exploiting shared neural activity for latent dynamics remains underexplored.

Purpose of the Study:

• To develop a novel meta-learning approach for inferring latent dynamics from neural activity across related tasks.
• To address limitations of single-dataset approaches in handling recording heterogeneities.
• To enable rapid learning of latent dynamics from new neural recordings.

Main Methods:

• Proposed a meta-learning framework to model a family of related solutions for similar tasks.
• Captured cross-recording variabilities within a low-dimensional manifold.
• Utilized task-related neural activity from trained animals.

Main Results:

• Demonstrated efficacy on few-shot reconstruction and forecasting of synthetic dynamical systems.
• Validated the approach on neural recordings from motor cortex during arm reaching tasks.
• Showcased the ability to rapidly learn latent dynamics from new recordings.

Conclusions:

• The proposed meta-learning approach effectively parametrizes and learns latent dynamics across related tasks.
• This method facilitates rapid adaptation and generalization in neural systems and machine learning models.
• It offers a promising direction for analyzing complex neural data with inherent variability.