The Information Processing Group’s research in neural engineering has involved two primary problems:
Movement intent decoders allow amputees to control prostheses in a natural way by interpreting motor bioelectrical signals, giving them the chance to perform day-to-day tasks. Such systems have to overcome a number of challenges before they can become practical. These challenges include understanding the complexity of the underlying biochemical systems, the recursive nature of the human decision making process, the limited amount of data typically available for training and the time varying properties of the nervous system. We propose (a) to use systems trained using Markov Decision Process models to learn complex tasks using neural networks (b) to use shared controllers that combine the decoded information with a higher level goal to perform tasks, (c) to devise an adaptive framework for neural decoders base on semi-supervised learning. Such neural prostheses should behave like actual arms and will enable amputees to regain the ability to perform most of the tasks that the original limbs could perform. This research not only has a compelling social impact, but also will contribute to the data science community.
Individuals with lower limb paralysis retain the ability to plan and initiate gait within the central nervous system but, due to conditions such as spinal cord injury or stroke, are unable to transmit those directives to the peripheral nervous system and the muscles of the lower limb. Although restoration of stance and gait by electrical stimulation of either peripheral nerves or lower limb muscles has a long history, such approaches have not reached widespread acceptance or use, which we propose is due to reliance on classical engineering approaches to solve a biological problem. In contrast, we propose to use a biologically-inspired, data-driven approach to address this deficiency of the biological system. To that end, we propose developing and evaluating the use of machine learning methods, such as the convolution neural network models that were inspired by the organization of primary visual cortex, to control activation of muscles via asynchronous intrafascicular stimulation peripheral nerves. Specifically, the goal of this project is to develop and evaluate methods to restore natural, coordinated, and graceful gait in an animal model of paralysis, an anesthetized feline. In particular, we have the goals to (a) develop, characterize, and evaluate advanced controllers of joint angle and joint torque production of a single joint in only a single direction, to allow comparison of the data-driven model to our earlier classical controls methods; (b) develop, characterize, and evaluate advanced controllers of joint angle and joint torque production of a single joint in both directions, to elucidate methods used by the advanced controller's solution to the underconstrained problem of agonist-antagonist muscle pair control; (c) develop, characterize, and evaluate advanced controllers of joint angle and joint torque production of multiple joints in oth directions, to elucidate methods used by the advanced controller's solution to the competing-goals problem of biarticular muscle control and, (d) recreate natural, coordinated, and graceful gait by use of the advanced controllers arising from the first three goals, and demonstrate this result in an anesthetized cat walking on readmill. The result of the last item, an advanced gait controller, will be evaluated under conditions that provide a clear measure of performance and scenarios that represent activities of daily living of those with paralysis. Our approach to restoring naturalistic gait is expected to provide fundamental information to the functional neuromuscular stimulation community as to more effective methods for restoration of naturalistic gait.