Learning Hierarchical Models for Motor Control
Research Area:Biomedical and Biologically Motivated Technical Applications
Researchers:Alessandro Salatiello; Martin A. Giese;
Collaborators:Neville Hogan; Tamar Flash; Dagmar Sternad
Proposed start date:2018-05-01
Proposed end date:2021-08-02
There is strong evidence that the animal Motor Control System is hierarchically organized into highly-interacting specialized subnetworks. Such distributed control architecture guarantees remarkable robustness, flexibility and dexterity, which roboticists have been striving to achieve.
Unsurprisingly, the identification of such architecture is a complex problem, which requires a joint effort of experimental and theoretical neuroscientists. A typical scenario in this research area is the one in which electrophysiologists collect data at only one level of the hierarchy (e.g. Primary Motor Cortex or Neuromuscular Interface) and they would like to uncover the underlying structure that generated the recorded pattern of activation. In our lab, we develop algorithms that help to identify such hierarchical structures automatically from available neural and behavioral data.
As one approach, we combine methods from System Identification Theory (e.g. Unscented Kalman Smoother) and Machine Learning (e.g. Dynamical Neural Networks and Structure Learning) to automatically identify the hierarchy of neural networks underlying the generation of movements, e.g. in the spinal chord. Present work focuses on finding robust algorithms and the analysis of the uniqueness of the solutions of the underlying system identification problem.
Funding: CRCNS US-German-Israeli Collaborative Research Proposal: Hierarchical Coordination of Complex Actions, funded by the BMBF (FKZ: 01GQ1704).