Hierarchical Bayesian Optimization in Neurostimulation

Neuroprostheses interfacing with brain and spinal tissue are in rapid development and are being applied to clinical treatment of paralysis after spinal cord injury stroke and other neurological disorders.
We have developed a versatile intelligent neuroprosthetic agent, a Gaussian Process (GP)-based Bayesian Optimization (BO) approach. We have further laid the theoretical groundwork to cover larger input spaces with a hierarchical GP.

Within this project, Julien will design a fully-hierarchical GP-BO, whereby multi-dimensional GP (i.e., a space of complex, multi-electrode stimulation patterns) rely on models preliminarily fitted in lower dimension (single-electrode). Beyond our previous work, the additional concept proposed here is that knowledge learnt during searches performed in the high-dimensional space is propagated back to the base low-dimensional components to update the building blocks of the hierarchical structure.

Faculty Supervisor:

Numa Dancause

Student:

Partner:

École polytechnique fédérale de Lausanne

Discipline:

Life Sciences

Sector:

Education

University:

Université de Montréal

Program: