Generative Decoding of Intracortical Neuronal Signals for Online Control of Robotic Arm to Intercept Moving Objects

Abstract

Abstract Brain-machine interfaces (BMIs) are being developed for translating neural signals to control an external device such as a robotic arm. The prevalent decoding strategy of current BMIs is the continuously conversion of neural spike trains to motor variables via a discriminative algorithm, but this typically results in un-natural and slow robotic motion. Recently, we implemented a BMI under feedforward control for interception of moving targets, though the decoding accuracy was insufficient for accurate online control. To improve the BMI performance, a generative algorithm called Latent Factor Analysis via Dynamical Systems (LFADS) [1], was applied to data pre-processing for de-noising. The results indicate potential advantages of feedforward control with generative decoding in improving BMI design.