Skip to content Skip to navigation
Qualifying Exam
5/7/2019 03:00 pm
CBIM 22

Brain-informed learning models of movement primitives for neurorehabilitation robots

Neelesh Kumar, Department of Computer Science

Examination Committee: Prof. Konstantinos Michmizos (Chair), Prof. Tomasz Imielinski, Prof. Ahmed Elgammal, Prof. Srinivas Narayana

Abstract

Neurorehabilitation robotics is the epitome of a challenging interaction between robots and humans, where pathophysiological movements need to be first accurately and precisely identified and then steered towards normalcy. In this talk, I will present our first fruits in exploring the fascinating possibility that a neural network, when designed and trained upon the brain's constraints, has enough discriminative power to predict brain states associated with movement primitives. We have initially focused on the intent of the user to move (2-classes), the reaction time (3-classes), and the direction of the movement (4-classes). Specifically, I will propose a new 3-dimensional Convolutional Neural Network architecture that accounts for the brain regions associated with the respective task. I will demonstrate how, for the first time, a model can learn to predict movement primitives using electroencephalography (EEG), in an experimental paradigm that we designed and ran in the Lab. In addition to establishing the effectiveness of 3D convolutions in extracting task-discriminative spatiotemporal features from EEG, our model outperformed the state-of-the-art methods, demonstrating above 9% improvement in the 4-classes task. By developing data-rich brain-constrained deep models, I seek to objectify and subsequently optimize therapy provided by robots as they interact with the intrinsic variability of each brain’s dysfunction and steer it towards normalcy.

(This research is supported through Grant K12HD093427 from the National Center for Medical Rehabilitation Research, NIH/NICHD.)