Recurrent neural network modeling of hierarchical motor control and analysis
Current theories on biological motor control mostly concern with how the brain as a whole performs sensory-motor processing, with little reference to how such functions are distributed across the hierarchical structure. Here, we study the fundamental design principle of biological motor control hierarchy, with a special focus on the lowest-level-controllers (LLC), such as the primary motor cortex (M1) and spinal cord central pattern generators (CPG). As a component of a hierarchical structure, a LLC receives command signals from higher-level-controllers and controls the body movement accordingly. Our study shows that an LLC should exhibit stability and memoryless properties in order for the command signal to be properly reflected in the body movement. These conditions imply that when command input is held fixed over time, the body state should converge to a stable attractor that is uniquely determined by the input. It also means that the mapping from command to body movement is unambiguous. Experiments confirm that biological motor systems indeed employ attractor dynamics. Graziano et all [1] showed that micro-stimulation of M1 drives limbs toward a unique posture regardless of previous movement history. The result implies that M1 has point-attractor dynamics. On the other hand, most CPGs in the spinal cord (e.g. locomotive CPG) have periodic movement patterns, which are limit-cycle attractors. By exploiting the attractor property, we calculate the true complexity (dimensionality) of the LLC mapping problem (It is roughly twice the dimensionality of the attractor space). We also obtain an efficient training method for approximating the optimal LLC that is otherwise a hopelessly high-dimensional learning problem. We train recurrent neural networks (RNN) to approximate desired LLCs. These RNN models unify multiple theories of motor control (optimal control, equilibrium control, coordinate translation, synergy) and provide a bridge that connects the theories and neurophysiology data. The RNN models reveal the following properties of LLC control: (1) Attractor dynamics of the LLC is the basic building block of all movement generation. (2) The combined dynamics of LLC-body is simpler than the pure body dynamics. (3) Due to the simplified dynamics, the higher-level-controllers can easily manipulate body movements without having to dealing with complex physical properties of muscles and joints. Our work questions the validity of the current view of motor control hierarchy, which assumes that different levels of hierarchy have distinct functional roles (decision making ->movement trajectory planning -> execution of motor plan). Valuable progress has been made in domains like sensory process by using feed-forward models with functionally distinct hierarchical processes. However, there is a growing need to account for dynamical systems with feedback and recurrent processing, which resists the same kind of functional decomposition. For example, controlling a movement trajectory does not only depend on a high-level-command sequence, but also the dynamic properties of LLC-body. A useful alternative is to consider the design principle of the whole hierarchy as successively simplifying the body dynamics, so that higher-level-controllers can achieve abstract motor goals without dealing with the complexity of body dynamics.
References
1. Graziano, et al, 2002. Neuron 34, pp. 841'851
Conference:
Computational and systems
neuroscience 2009, Salt Lake City, UT, United States, 26 Feb - 3 Mar, 2009.
Presentation Type:
Poster Presentation
Topic:
Poster Presentations
Citation:
(2009). Recurrent neural network modeling of hierarchical motor control and analysis.
Front. Syst. Neurosci.
Conference Abstract:
Computational and systems
neuroscience 2009.
doi: 10.3389/conf.neuro.06.2009.03.262
Copyright:
The abstracts in this collection have not been subject to any Frontiers peer review or checks, and are not endorsed by Frontiers.
They are made available through the Frontiers publishing platform as a service to conference organizers and presenters.
The copyright in the individual abstracts is owned by the author of each abstract or his/her employer unless otherwise stated.
Each abstract, as well as the collection of abstracts, are published under a Creative Commons CC-BY 4.0 (attribution) licence (https://creativecommons.org/licenses/by/4.0/) and may thus be reproduced, translated, adapted and be the subject of derivative works provided the authors and Frontiers are attributed.
For Frontiers’ terms and conditions please see https://www.frontiersin.org/legal/terms-and-conditions.
Received:
04 Feb 2009;
Published Online:
04 Feb 2009.