Control

Our understanding of motor control has long been limited by models that oversimplify arm and hand control. Capturing real behaviour requires models that account for muscle dynamics, the delays and noise inherent in biological systems, sensory feedback arriving during movement, and the complementary roles of different brain regions.

We build neural network models of closed-loop control that integrate all of these factors, using tools we developed such as MotorNet, a Python toolbox for training networks to control biomechanically realistic limbs. Linking the strategies of these artificial networks to the brain has already proven fruitful: our closed-loop network models predicted how sensory expectations should reshape neural population dynamics, which we then confirmed in recordings from motor circuits. We are now using this approach to understand how multiple brain regions divide up the work of control, and how control policies are learned, retained, and re-used across contexts.

Neural network model trained perform sequential reaches by controlling simulated muscles (related paper)