Contrasting action and posture coding with hierarchical deep neural network models of proprioception

Area of Science:

• Neuroscience
• Robotics
• Biomechanics

Background:

• Biological motor control relies on proprioceptive feedback for adaptability and efficiency.
• Proprioception typically represents body state, but its role in higher-level functions is less understood.

Purpose of the Study:

• To investigate the hypothesis that proprioception is critical for action recognition.
• To isolate and study the function of proprioception using a task-driven modeling approach.

Main Methods:

• Generated synthetic 3D arm trajectories and corresponding muscle spindle activity data.
• Trained hierarchical artificial neural networks for trajectory decoding and action recognition tasks.
• Compared neural network unit tuning properties to primate brain data.

Main Results:

• Artificial neural networks successfully decoded both arm trajectories and action identities from proprioceptive data.
• Network units exhibited tuning properties mirroring those in the primate somatosensory cortex and brainstem.
• Uniformly distributed directional selective units were found exclusively in action-recognition-trained models.

Conclusions:

• Proprioceptive encoding is associated with higher-level cognitive functions like action recognition.
• These findings offer new, testable hypotheses on proprioception's role in adaptive motor control.