Echo State Networks for Motor Control of Human ECoG Neuroprosthetics

Towards non-invasive neuroprosthetic systems for motor control, electrocorticogram (ECoG) recordings provide an intermediate step from microwire single neuron recordings to electroencephalograms. Preprocessing modalities that emphasize amplitude modulation and temporal power in the ECoG are employed to build human brain-machine interfaces, which have been previously shown to modulate with hand behavior by means of linear filters, though, with limited accuracy. We improve the online decoding performance of the amplitude modulation across the recording spectra by employing echo state networks (ESNs) which provide nonlinear mappings without compromising training complexity and filter order compared to basic linear filters and other neural networks. Preliminary results show an increase of 15% in the average correlation of ESN outputs with actual hand trajectories compared to linear mappings.