Transfer Learning In Brain-Computer Interfaces

Brain-computer interface(BCI)provides a direct communication pathway between the brain and a computer.In other words,it enables a user to control external devices by using brain signals,which can be applied in biology,medical treatment,and engineering.Electroencephalogram(EEG)may be the most popular BCI input signal,which is collected from the scalp.BCI recognizes the brain activity patterns and the user's intentions by decoding EEG,and translates them to control signals for external devices.However,due to individual differences,i.e.,different subjects have different neural responses to even the same stimulus,it is very difficult to build a generic model whose parameters fit all subjects.So,it usually needs a long calibration process to collect labeled subject-specific data for each individual subject,which is not user-friendly.Transfer learning,which transfers information from related domains to a target domain,is a promising approach for this problem.More specifically,transfer learning reduces the distribution gaps between the auxiliary subjects(source subjects)and the new subject(target subject),and leverages data or knowledge from source subjects to improve the learning performance for a target subject,so that the calibration process can be shorten or even eliminated.This dissertation focuses on transfer learning in BCIs.We proceed from algorithms to data,and from homogeneous to heterogeneous scenarios,and propose several approaches for transfer learning.In homogeneous scenarios,transfer learning is first incorporated into spatial filtering algorithms.Then improvements over a state-of-the-art transfer learning algorithm are proposed.Afterwards,a data alignment approach is proposed to align EEG trials,which could be used for different algorithms and models.In heterogeneous scenarios,the problem of heterogeneous label spaces is studied and a label space alignment approach is proposed.The major contributions of this dissertation are:1.For addressing the problem of incorporating transfer learning into common spatial pattern,which is the most widely used spatial filtering algorithm in BCI,an instance-based approach is proposed.More specifically,it reduces the distribution discrepancies between the source and target subjects by re-weighting the raw signals or the tangent space vectors of covariance matrices,and then uses the reweighted sample covariance matrices from source subjects to improve the computation of common spatial pattern filters for the target subject.Experimental results show the performance on two motor imagery datasets was improved by 9.34%,which is greater than that of several existing relevant approaches in literature.2.Two approaches are proposed to improve the transfer learning algorithms designed in a Riemannian framework.The first one is channel selection,which reduces the dimension of covariance matrices,and hence makes the Riemannian space computations more accurate and efficient.The second is trial selection,which resamples trials from source subjects using a Riemannian distance-based clustering,such that they become more consistent with those of the target subject.Experimental results show that both approaches enhance the performance of one of the state-of-the-art transfer learning algorithms.More specifically,the first approach improved the average accuracy across all subjects by 15.39%,and made the algorithm 30 times faster.And the average accuracy was further improved by from2.71% to 4.54% when performing both approaches.3.An unsupervised data alignment approach is proposed to directly align the EEG trials from different subjects.For each subject,it uses the mean covariance matrix as a reference to transform all trials,such that the mean covariance matrix of the transformed trials becomes an identify matrix.The data of different subjects become more similar as their covariance matrices are aligned.Experimental results show that this data alignment approach significantly outperforms state-of-the-art transfer learning algorithms.More specifically,the performance on motor imagery datasets was improved by around 11%,and that on P300 dataset was improved by around 9%.4.A label alignment approach is proposed for heterogeneous label spaces transfer learning.More specifically,it first matches each source subject label with a target subject label,then moves the per-class covariance matrices of the source subject to re-center them at the corresponding class means of the target subject,and updates the source labels to the matched target labels.Experimental results show that this label alignment approach is effective in multiple scenarios of heterogeneous label spaces,and can be integrated with other transfer learning approaches to achieve even better performance.In conclusion,this dissertation proposes several transfer learning approaches to address the problems of cross-subject,cross-headset and cross-task for brain-computer interfaces.The experimental results demonstrated the effectiveness and efficiency of the proposed approaches,which can be used to reduce the calibration effort of brain-computer interfaces.