Inference and Decoding of Motor Cortex Low-Dimensional Dynamics via Latent State-Space Models

Area of Science:

• Neuroscience
• Computational Neuroscience
• Motor Control

Background:

• Motor cortex exhibits low-dimensional collective dynamics in neuronal spiking activity during behavior.
• These dynamics represent coordinated modes of neural activity that are crucial for motor control.

Purpose of the Study:

• To demonstrate that unsupervised latent state-space models can reveal low-dimensional motor cortex dynamics.
• To compare the accuracy of movement kinematics decoding from these dynamics versus direct decoding from neuronal ensembles.

Main Methods:

• Recorded single neuron ensembles in nonhuman primate motor cortex (PMv, PMd, M1) during 3-D reach-to-grasp actions.
• Estimated low-dimensional dynamics using Poisson linear dynamic system (PLDS) models.
• Implemented decoding via point process and Kalman filters, comparing latent state-space models with predictive subsampling.

Main Results:

• Inferred low-dimensional motor cortex dynamics preserve naturalistic reach kinematics information.
• Decoding from latent state-space models achieved accuracy comparable to or better than direct decoding from the entire neuronal ensemble.
• Decoding based on unsupervised PLDS models outperformed previous approaches, including predictive subsampling.

Conclusions:

• Low-dimensional dynamics inferred by unsupervised models effectively capture essential information for movement decoding.
• Unsupervised PLDS models offer a powerful and potentially superior method for decoding motor kinematics from neural population activity.
• These findings highlight the significance of collective neural dynamics in motor cortex function and decoding.