Towards autonomous neuroprosthetic control using Hebbian reinforcement learning.

Area of Science:

• Neuroscience
• Robotics
• Machine Learning

Background:

• Neuroprosthetic systems require sophisticated controllers to translate neural signals into prosthetic actions.
• Current systems often lack autonomous adaptation capabilities, limiting user experience and performance.
• Developing controllers that learn and adapt based on user feedback is crucial for advancing neuroprosthetics.

Purpose of the Study:

• To design an adaptive neuroprosthetic controller capable of learning neural state-to-action mappings.
• To enable automatic adaptation using only binary evaluative feedback.
• To investigate the convergence and performance of the proposed adaptive controller.

Main Methods:

• Utilized Hebbian reinforcement learning (HRL) within a connectionist network for controller design.
• Combined supervised and reinforcement learning principles for efficient and generalizable control.
• Evaluated convergence properties through closed-loop simulations and open-loop simulations with primate neural data.

Main Results:

• The HRL controller demonstrated proficiency in classification and regression tasks.
• Achieved rapid convergence to an effective control policy with robust performance.
• Successfully halted adaptation upon reaching satisfactory performance and resumed adaptation when input changed.

Conclusions:

• The HRL control algorithm offers an efficient pathway for autonomous adaptation in neuroprosthetic systems.
• Enables users to guide controller behavior via simple evaluative feedback.
• Potential to significantly enhance user interaction and system adaptability in neuroprosthetics.