Cortical motor intention decoding on an analog co-processor with fast training for non-stationary data

Abstract

This paper presents a low power hardware implementation of a motor intention decoder used in intra-cortical Brain Machine Interfaces. It offers two specific advantages over current state of the art decoders. Firstly, the decoding is done on an analog co-processor instead of a personal computer thereby reducing both the power consumption and size of the overall system. Secondly, the co-processor employs a randomized neural network — extreme learning machine (ELM), which is as quick to train as the linear decoders while being adept at capturing the complex non-linear mappings between the neural activity and the intended movements. Results show an average 10% improvement in decoding accuracy over linear discriminant analysis in non-stationary datasets.