Zero-training Motor-Imagery Brain-computer Interface

Brain-computer interfaces(BCIs)provide a channel to communicate directly with the outside world,bypassing the neural system and muscle and have demonstrated their potentials across virous applications.Due to the nonstationary character of EEG,subjects have to undertake boring and time-consuming calibration sessions before use,which prevents BCI from going to real applications.The main work of this paper is to take use of other subjects’ motor-imagery(MI)data to solve this problem in feature extraction,feature fusion and classifier design.Therefore,the common and different parts of MI data,including how to alleviate the discrepancy across subjects were investigated.Common and stationary features were extracted and three methods of reducing the differences between subjects were proposed.Finally,zero-training MI BCI were realized exploring other subjects’ data.The work and innovations of this paper are as follows:1)Conventional common spatial pattern(CSP)cannot work effectively under zero-sample setting.The common part of CSP filters across subjects were investigated and then common stationary spatial filters across subjects were extracted by a novel clustering method.It is proved that the feature-extracting method had a good performance.After that,algorithm called SemiTr SVM was proposed combining transfer learning and semi-supervised learning.It utilized source data to construct a classifier and experiments illustrated it was very high-performance for target subjects.2)In order to reduce the difference between the source and the target domain,and make the models trained in the source domain work well for target domain,kernel embedding for probability in the reproducing kernel Hilbert space was applied,and two methods to measure the discrepancy between source and target distribution were proposed,called CDD and JDD.Experiments demonstrated these two methods can well measure the discrepancy between source and target distribution.The joint distribution of the target and the source domain were made closer.Algorithms called CDDSVM and JDMSVM were proposed,which realized zero-training for BCI.Comparisons demonstrated the outstanding performances of our algorithms.3)Applying the kernel embedding theory,an end-to-end deep neural network was proposed,based on multi-level joint distribution adaptation.This network combined the characteristics of time,frequency,and space domain of motion imagery EEG signal,and effectively integrated the data of source and target subjects into a common space.Experiments demonstrated the proposed deep network was very high-performance across subjects.To summarize,after investigating the similarities and differences of different subjects’ for motor imagery,this paper proposes a series of methods for cross-subjects BCI,including the aspects of feature extraction and adaptation,adaptive classifier and deep neural network.Finally,zero training motor imagery BCI is realized.