Simulation of two methods in co-adaptive control for brain-machine interfaces

Abstract

Simulation of two methods in co-adaptive control for brain-machine interfaces: cursorGoal (developed by the Shenoy Lab @ Stanford EE and the Carmena Lab @ UC Berkeley EE), and Joint RSE (developed by the Neural Signal Processing Laboratory, www.nsplab.org). The healthy volunteer represents a sensorimotor neural control network. His arm movements are captured with the Microsoft Kinect and used to drive simulated neural activity (not shown) from a point process model of primary motor cortex. This neural activity determines movements of the on-screen cursor through a brain-machine interface (BMI) algorithm. In all trials, the cursor begins at a random point on the outer circle, and the user attempts to adjust his arm movements to bring the cursor to the inner circle (target) for a specified hold period of 0.5 sec. Maximum allowed trial time is 3 sec. In the training trials of these simulations, the various BMI algorithms must both learn neural signal parameters and decode arm movement. In test trials, the neural signal parameters are fixed, and both methods use an identical filter formulation (Eden, 2004 Neural Computation) with a random walk state equation to drive arm movements. cursorGoal and Joint RSE differ markedly in the way visual feedback to the user (cursor movement) is determined during training trials, as well as in the procedure for learning neural signal parameters. The related manuscript delineates the relative contributions of these algorithmic variations to the differing performance of these co-adaptive BMI control methods, where Joint RSE consistently and substantially outperforms cursorGoal.