An Open Source Simulator for Multi Degree-Of-Freedom Brain-Machine Interfaces

点击翻译按钮获取中文摘要

PROJECT SUMMARY For millions with movement disorders including paralysis and ALS, intracortical brain-machine interfaces (BMIs) are an emerging technology that aims to restore lost motor function and communication. The main component of a BMI is a decoder algorithm that translates neural activity from motor areas of the brain into the kinematics of a prosthetic device. Due to the complexity of these systems, which includes the BMI user interacting with the decoded kinematics in a closed-feedback loop, current technology requires expensive and invasive experiments to design, optimize, and validate decoder algorithms. The need for such experiments (1) results in slow develop- ment and evaluation of decoder algorithms, and (2) limits the scope of people who can work on these problems to a small group of nonhuman primate and clinical trial labs. As a consequence, BMIs have remained in pilot clinical trials since their first reported demonstration in 2004. We propose a new open-source simulator for multiple degree-of-freedom (DOF) BMI systems. The goals of this simulator are to (1) reduce the time it takes to evaluate and optimize BMI algorithms from months to minutes, and (2) significantly expand the community of researchers who develop testable algorithms for BMIs. To build the simulator, we propose neural encoding models that generate synthetic motor cortical activity for multiple DOF tasks. This is possible because neural population activity is relatively low-dimensional and has dynamics, which can be learned via recurrent neural networks (RNNs). We build our neural simulators using data collected from human clinical trials during point-to-point multi-DOF reaches. We also propose to develop new models of human controllers. This solves an important problem in BMIs: users learn new control strategies when controlling a particular BMI decoder algorithm. Our simulator uses deep imitation and reinforcement learning to solve this problem. It is constrained through imitation learning to perform actions like a human. It is optimized through reinforcement learning to explore new strategies – under the constraint of being human-like – to optimally control the BMI. Together, we expect these innovations will result in a purely software simulator that accurately predicts BMI performance and enables design and optimization. This tool will be open-sourced and available to all, enabling widespread development of BMIs.