| Brain-computer interface(BCI)is a technology that provides a direct interaction pathway between the brain and external devices without relying on the peripheral nervous system and muscles.It collects brain signals and converts them into computer commands to control external devices.Due to the fact that electroencephalography(EEG)is highly noisy and varies by subject,traditional BCIs usually require full calibration to adapt to the user’s brain patterns before use.However,the calibration process is tedious and timeconsuming,making BCI systems impractical for everyday use.To completely eliminate or dramatically shorten the calibration process and substantially improve the convenience of the use of BCIs,we propose a variety of approaches for cross-subject EEG analyses,and further establish online P300 BCI systems with zero or shortened calibration based on these approaches.Experiments on 200 subjects demonstrate the effectiveness of the proposed approaches.First,we propose an invariant pattern learning approach based on a CNN and big EEG data,and build a subject-independent BCI online system with zero calibration.Specifically,we collected EEG data from 200 subjects in a BCI-based spelling task using two different types of amplifiers.A CNN was trained using EEG data from 150 of these subjects,allowing it to extract subject-independent features and make predictions for new users.Based on this approach,a subject-independent online BCI system with zero calibration was developed.The offline analysis on independent test sets containing EEG data from the other 50 subjects showed that almost all subjects obtained significant cross-subject and cross-amplifier effects,with an average accuracy above 80%.Furthermore,more than half of the subjects achieved accuracies above 85%.Twenty subjects participated in an online experiment,and the average accuracy reached 89.38%.These results indicate that our method is effective for building a subject-independent BCI.Second,based on the above subject-independent model,we propose two approaches for model’s online learning based on unlabeled data collected during the user’s online operation,and establish online BCI systems with zero calibration and with the capability of online learning.The first approach is online learning based on an unsupervised domain confusion algorithm.In the model retraining phase,the algorithm optimizes the classification loss on a large-scale dataset,as well as enhances the invariance between the distribution of unlabeled data from the target user and the distribution of data from the large-scale dataset by introducing a maximum mean discrepancy(MMD)loss function.The experimental results showed that on independent test sets containing data from 50 subjects,this approach achieved an average accuracy above 85%,which was improved compared with the results of the subject-independent method.In addition,the proportion of subjects with high accuracies was increased as well.The second approach is online learning based on a semisupervised self-training algorithm.In the model retraining phase,this approach iteratively assigns predicted labels to unlabeled data collected during the user’s online operation,selects data with high decision-making confidence,and then retrains the model with these data along with the predicted labels.The offline analysis showed that an average accuracy of 92.13% was achieved on the independent test set containing data from 50 subjects.In addition,86.00% of the subjects achieved accuracies above 85%.Twenty subjects participated in the online experiment,and the average accuracy reached 94.00%.These results demonstrate that without calibration,the proposed approach can achieve comparable performance to traditional methods with full calibration.Third,to solve the issue that zero-calibration approaches are not perfectly suitable for a portion of new users in the early stage of their operation,we propose a supervised transfer learning-based approach to update the subject-independent model,and establish an online BCI system with shortened calibration.Specifically,we collected a small quantity of labeled data before the user’s online operation,and used these data to fine-tune the model based on the subject-independent model.Based on this approach,we developed an online BCI system with shortened calibration.The offline analysis showed that with approximately 1 min of calibration data collection,this approach achieved an average accuracy of 90.82% on independent test sets containing data from 50 subjects.Twenty subjects participated in the online experiment,and an average accuracy of 93.50% was achieved.These results demonstrate that this approach can achieve comparable performance to traditional systems with full calibration throughout the user’s online operation,while the calibration process is dramatically shortened.Last,we present a method for EEG channel selection across subjects for building fewchannel BCI systems with zero/shortened calibration.Specifically,we selected a fraction of EEG channels corresponding to the largest weights obtained from the spatial filters in the subject-independent full-channel models,and subsequently rebuilt the models with these channels.Experimental results showed that with zero or shortened calibration,average accuracies of around or above 80% were achieved with only 5–10 EEG channels,which demonstrates the effectiveness of our method. |
