Deep networks for motor control functions | JoVE Visualize

Area of Science:

Motor control and learning
Computational neuroscience
Machine learning applications in biology

Background:

Conventional motor control models use dynamical representations (forward/inverse models) that focus on spatial states, not temporal command generation.
Generating time-varying motor commands requires integrating control policies forward, which can be computationally intensive and less intuitive for neural implementation.

Purpose of the Study:

To introduce and evaluate a novel functional approach to motor control that directly represents time-varying commands and state trajectories across both space and time.
To leverage advances in machine learning, specifically deep networks, to implement this functional representation and overcome challenges associated with increased parameters.

Main Methods:

Utilized a stacked autoencoder (deep network) to create a function that maps initial and target states to optimal time-varying command and state trajectories.
Trained the deep network on a non-linear limb model, including scenarios with varying force fields, to generate accurate movement profiles.
Demonstrated the network's ability to learn through trial and error and optimize a cost objective, mimicking aspects of motor babbling and learning.

Main Results:

The deep network successfully generated accurate temporal profiles of optimal commands and state trajectories for point-to-point reaches.
The functional approach demonstrated adaptability to changing dynamics, such as varying force fields.
The network exhibited self-learning capabilities and the ability to optimize movement costs.

Conclusions:

A functional, deep learning-based approach offers a powerful alternative to traditional dynamical models for understanding motor control and learning.
This method provides a representation across both space and time, potentially offering new insights into the neural mechanisms underlying motor control.
The findings suggest that complex motor behaviors, including learning and optimization, can be modeled effectively using deep neural networks.