Neural Decoding Based on Active Learning for Intracortical Brain-Machine Interfaces

Abstract

In intracortical brain-machine interfaces (iBMIs), it is time-consuming and expensive to label the large number of unlabeled samples. In this paper, three greedy sampling active learning algorithms, named denoised greedy sampling on the inputs (DGSx), denoised greedy sampling on the outputs (DGSy) and denoised improved greedy sampling (DiGS), were proposed to solve the problem of labeling samples. In iBMIs, in order to reduce the influence of abnormal points in the raw data, Oneclass-SVM was assumed to denoise to improve the performance of original greedy sampling algorithms and achieve stable or robust decoding. Compared with Query by Committee (QBC) and Uncertainty Sampling (US), the proposed approaches achieved higher accuracies. The efficiency of proposed approaches was demonstrated by the experiment with rhesus' electrophysiological signals in iBMIs.