Deep networks for motor control functions

Max Berniker, Konrad P. KördingFrontiers Comput. Neurosci 2015 · DOI

Abstract

Traditional motor control frameworks rely on dynamical models that integrate instantaneous states over time, a process that is computationally demanding and assumes a specific temporal representation in the brain. This paper proposes a functional alternative: representing entire movements as single mappings from goals to space-time trajectories. Using deep stacked autoencoders, the authors demonstrate that neural networks can learn to output complete motor command profiles and state estimates. These networks outperform linear baselines, compensate for complex force fields, and can bootstrap their own learning through "motor babble" without requiring optimal teachers. This approach suggests a neural architecture where time is explicitly represented across distributed populations.

Introduction: From Dynamics to Functions

The standard framework for motor control is rooted in robotics and Newtonian physics, relying on dynamical models—forward and inverse models that represent the body's state at a specific instant. To generate a movement, the nervous system must integrate these models forward in time. While mathematically sound, this approach is computationally intensive and assumes the brain represents "space" but not "time" explicitly.

The authors ask: Could the nervous system instead represent an entire movement as a single function? Instead of calculating the next millisecond of a reach based on the current one, could a network take the "start" and "end" points as inputs and immediately output the entire time-varying command profile? This "functional" approach represents the trajectory across both space and time simultaneously.

Conceptual comparison between dynamical motor control (A), algebraic models (B), and the proposed trajectory function approach (C), alongside command (D) and state (E) profiles.

Methods: Learning the Movement Manifold

The authors reframe motor control as a function approximation problem: $Y = F(X)$, where $X$ represents the task parameters (initial state, target, and force field conditions) and $Y$ is the vectorized trajectory—a sequence of all states and motor commands over the duration of the movement.

To handle the high dimensionality of $Y$, which contains hundreds of time-sampled points, they employ Deep Stacked Autoencoders. The training process involves three key stages:

  1. Unsupervised Pre-training: The network first learns to compress high-dimensional trajectory data into a low-dimensional feature bottleneck ($z$).
  2. Supervised Mapping: A shallow network is trained to map the task inputs ($X$) to these low-dimensional features ($z$).
  3. Refinement: The layers are "unrolled" and fine-tuned to produce the final trajectory output.

This deep learning approach was compared against a linear baseline using Principal Component Analysis (PCA) to determine if non-linear architectures are necessary for capturing the complexities of motor control.

Experimental Results

The study utilized simulated data from a two-link arm model, incorporating optimal trajectories and various force field conditions (no force, clockwise, and counter-clockwise curl fields).

Superior Accuracy and Force Field Compensation

The deep network achieved a validation error of 0.004, an order of magnitude better than the PCA baseline (0.039). While the linear PCA model estimated nearly straight lines regardless of the environment, the deep network accurately predicted the curved trajectories required to compensate for curl fields. This suggests that optimal trajectories reside on a non-linear manifold that deep networks are uniquely suited to identify.

Bootstrapping via Motor Babble

A significant finding is the network's ability to learn without an "optimal teacher." By starting with random "motor babbling" torques, the network observed the resulting limb trajectories and learned the relationship between commands and outcomes. After three rounds of iterative self-generation, the network could reach targets with an average Euclidean error of only 0.357 cm, eventually refining commands to minimize metabolic costs.

Discussion and Neural Plausibility

Unlike a controller that reacts moment-to-moment, this functional network produces a "feedforward" plan for the entire movement duration at once. The authors suggest this aligns with evidence of "spatially distributed" representations in the brain, where different groups of neurons might encode the temporal profile of a movement at distinct points in time.

While the current model uses a fixed movement duration, future work may incorporate temporal basis functions (like wavelets) to allow for temporal scaling. Additionally, probabilistic extensions using Deep Boltzmann Machines could create bidirectional models that function as both forward and inverse models simultaneously.

Summary updated

BibTeX
@article{berniker2015deep,
  author = {Max Berniker and Konrad P. Körding},
  title = {{Deep networks for motor control functions}},
  journal = {Frontiers Comput. Neurosci},
  year = {2015},
  doi = {10.3389/fncom.2015.00032},
}
Figure