An in-silico framework for modeling optimal control of neural systems

Area of Science:

• Neuroscience
• Control Theory
• Machine Learning

Background:

• Brain-machine interfaces (BMIs) offer potential for restoring lost function but face challenges with complex neural systems.
• Classical control theory is limited in designing effective control laws for large-scale neural networks.

Purpose of the Study:

• To propose a scalable, data-driven, unified framework for studying brain-machine-environment interaction.
• To develop a methodology applicable to designing control algorithms for neural prostheses.

Main Methods:

• A unified approach defines neural systems, prostheses, and environments using differential equations with learnable parameters, reducing to recurrent neural networks.
• Three training methods are explored: direct optimization, oracle-based learning, and reinforcement learning, adapted to varying system knowledge.

Main Results:

• The framework was applied to an in-silico neural system for particle stabilization and pole balancing tasks.
• A prosthetic controller was trained to restore neural system behavior under simulated sensor and motor impairments.

Conclusions:

• The proposed framework facilitates rapid, flexible synthesis of control algorithms for neural prostheses, potentially reducing in-vivo testing.
• Implications for optimizing prosthetic sensor and actuator placement are discussed.