BMI Progress Reports 2021 // Prof. A. Mathis's Lab, Axel Bisi "Using task-driven deep neural networks to investigate the proprioceptive pathway"
Biological motor control is adaptive, efficient and is in part enabled by proprioception. How proprioception provides feedback to the sensorimotor system is not fully understood. We pursue a task-driven modeling approach that has provided important insights into other sensory systems. In order to do so, we first generated a large-scale dataset of human arm trajectories as the hand is tracing the alphabet in 3D space, then using a musculoskeletal model derived the muscle spindle firing rates during these movements. For this action recognition task, artificial neural networks were trained to classify the character identity from this proprioceptive stimuli. Network units showed directional tuning akin to neurons in the primate somatosensory cortex, and task learning yielded invariant tuning across a 3D workspace. We then used these task-driven models were to predict neural activity in primates. While not being explicitly trained to predict neural data, the models’ internal representations predicted neuron responses with high fidelity. Overall, our work links tuning properties in the proprioceptive system to the behavioral level, and explores neural coding with task-driven models in the proprioceptive pathway.