Model Predictive Control on the Neural Manifold.

Area of Science:

• Computational Neuroscience
• Systems Neuroscience
• Control Theory

Background:

• Neural manifolds offer a theoretical framework for understanding neural population activity.
• Current methods for identifying neural manifolds are often correlational, limiting insights into circuit dynamics.
• Precise control over latent neural activity is crucial for investigating neural manifold structure and function.

Purpose of the Study:

• To simulate and evaluate methods for controlling the latent dynamics of a neural population within its manifold.
• To compare the efficacy of proportional-integral-derivative (PID) control and model predictive control (MPC) for trajectory following in latent space.
• To establish a framework for experimentally testing causal relationships between neural manifold dynamics and external stimuli.

Main Methods:

• Utilized a spiking neural network (SNN) to model neural circuit dynamics.
• Simulated closed-loop, dynamically generated sensory inputs to control latent activity.
• Applied and compared PID and MPC control strategies for trajectory-following tasks.
• Evaluated controller performance under partial observability and unknown noise conditions.

Main Results:

• Identified low-dimensional representations for both neural population activity (neural manifold) and visual stimuli.
• Both PID and MPC controllers showed some success in controlling latent dynamics.
• Model predictive control (MPC) consistently achieved more accurate control and required less hyperparameter tuning compared to PID.
• Demonstrated the application of MPC on neural manifolds using data-driven dynamics models.

Conclusions:

• Model predictive control (MPC) provides a powerful and accurate method for manipulating neural population dynamics within a manifold.
• The simulated framework enables experimental testing of causal links between manifold dynamics and sensory inputs.
• This approach advances the study of neural circuit function and the structure-activity relationships within neural manifolds.