Multi-source manifold feature transfer learning with domain selection for brain-computer interfaces☆
Introduction
A brain-computer interface (BCI) offers a direct communication pathway for users to interact with the environment, via their brain signals containing abundant information of the users’ cognitive state or intentions [1]. BCI has great research importance and application value in many fields, such as auxiliary control, entertainment and rehabilitation training [2]. Electroencephalogram (EEG) is the most widely used technology for BCI systems, because of its non-invasive nature and high temporal resolution. Different neurophysiological paradigms of EEG have been used to operate BCIs, such as motor imagery (MI), event-related potentials (ERPs) and steady-state visual evoked potentials (SSVEP) [3]. In MI tasks, the users imagine movements of their body parts (e.g., hands, feet, and tongue), without actually performing the movement and even without tensing the muscles, which causes the modulations of brain rhythms in the involved cortical areas. In ERP tasks, the user is stimulated by a majority of non-target stimuli but a few target stimuli, wherein a specific ERP pattern appears in the EEG response after the user perceives a target stimulus. However, EEG signals are notoriously difficult to analyze due to the low signal-to-noise ratio and significant subject-to-subject variations. Therefore, it is of great interest to be able to extract and recognize EEG features with machine learning methods with shallow and deep structures [4], [5], [6], [7], [8].
Despite several recent advances, most of the BCI systems are still faced with a major challenge of long calibration time. Different users have different neural responses to the same stimulus, and even the same user differs the response to the same stimulus, especially at different time/locations. Besides, calibrating the BCI system require sufficient labeled data to train the subject-specific BCI model, and thus the calibration phase is usually time-consuming. Transfer learning is helpful for BCIs to conquer the abovementioned challenges [9]. The knowledge in one or more source domains is fully utilized to help the classification in a target domain, which improves the calibration performance. And cross-subject classification has become one of the most popular directions where the generalization ability (of the classification model) can be enhanced by improving the cross-subject performance. Recently, there is a trend to utilize the covariance matrices of EEG trials, which are symmetric positive definite (SPD) and can be viewed as points on an SPD manifold, the covariance matrix can boost the role of Riemannian geometry in BCIs. The Riemannian alignment (RA) framework was proposed to align EEG covariance matrices from different source domains, which can quickly reduce the distribution difference between the source and target domains by changing the reference position on an SPD manifold [10]. However, the Riemannian alignment approaches are computationally expensive, and not compatible with the machine learning approaches defined in Euclidean space.
Transfer learning methods for BCI usually aim to solve single-to-single (STS) transfer problems. In practice, it is noted that a good source domain can help to obtain a high classification accuracy, even with a simple transfer learning algorithm. Therefore, the quality of the source domain is important. However, multiple source domains are factually available for EEG signal classification, such as labelled data from other subjects or other sessions used before. In multiple source domains cases, good source domains are more likely to be found. Due to the data expansion, in multi-source transfers, good source domains can reduce the negative transfer (NT) caused by bad source domains. Therefore, multi-source transfer learning usually achieves more stable and higher classification accuracy than STS transfer learning. Multi-source unsupervised domain adaptation is proved to be valuable in performance improvement [11], [12], [13].
Despite the progress mentioned above, there are some unresolved difficulties and challenges in the field. Firstly, traditional transfer learning methods usually seek to adapt marginal and conditional probability distributions [14], [15]. As an efficient way for EEG transfer, the Riemannian alignment approaches usually aligns distribution centers and cannot completely minimize the marginal probability distribution difference. If the aligned features can be transferred by traditional transfer learning methods further, marginal probability distribution still need to be adapted in the subspace. Moreover, the existing transfer learning methods for BCI are lack of information exploration between multi-source domains. It is necessary to propose a multi-source transfer learning framework with simple structure which can integrate traditional transfer learning method and Riemann alignment approach. Secondly, class imbalance often exists in many BCI scenarios. Most transfer learning methods ignore this issue by treating the classes as balanced across domains, or they only handle the bias in one domain [14], [16], and this may hinder the effectiveness of transfer learning. Therefore, how to handle the class imbalance situation in transfer learning for BCI is necessary. Thirdly, to overcome NT, many domain selection methods were proposed but most of them are based on mathematical similarity calculation [17], [18], which aims to explore the relationship between one source domain and the target domain, but lacks the exploration of the relationship between multiple source domains.
Different from the existing results, this paper presents a multi-source manifold feature transfer learning (MMFT) framework with simple structure and strong stability. More specifically, the distribution means of source and target domains are aligned on the SPD manifold, and then the tangent space features are extracted using tangent space mapping in the proposed MMFT. Next, the tangent space features are reconstructed to Grassmann manifold features via the Geodesic flow kernel (GFK) approach [18]. Finally, the classifier of MMFT is trained by adapting condition probability distribution in the structural risk minimization (SRM) function. Compared with existing multi-source transfer learning algorithms, the new method has several noteworthy advantages:
1) MMFT minimizes the marginal probability distribution on the Riemannian manifold, and then transfers the manifold features with conditional probability distribution adaptation. To the best of our knowledge, it’s the first attempt to align the two probability distributions separately in the field of EEG signal classification, where the calculation of their weights is avoided. Additionally, the MMFT has a simple structure with only three hyperparameters to be determined, and classification accuracy is not sensitive to the changes of these hyperparameters.
2) To better align conditional probability distributions for each pair of source and target domains, the MMFT utilizes a voting mechanism to transfer knowledge from source domains to the target domain individually, instead of transferring all source domains together.
3) To tackle the class imbalanced scenarios, a weighted MMFT algorithm is proposed which boosts the performance of MMFT for imbalanced training datasets.
4) To reduce the negative transfer in multi-source transfer learning, a label similarity analysis for MMFT (LSA-MMFT), is proposed for domain selection, which improves the classification accuracy and the time efficiency of MMFT.
The remainder of this paper is organized as follows. Section 2 introduces related work on Riemannian geometry, subspace adaptation and domain selection. Section 3 describes the details of the proposed MMFT, w-MMFT and LSA-MMFT. In Section 4, the experiments verify the performance of MMFT compared with several state-of-the art transfer learning approaches, and the performance of LSA-MMFT with two other state-of-the art domain selection approaches. Finally, the conclusion is given in Section 5.
Section snippets
Riemannian geometry
Riemannian geometry is the branch of mathematics that studies smoothly curved spaces that locally behave like Euclidean spaces. There are three main types of Riemannian manifolds, i.e., symmetric positive definite (SPD) manifold, Stiefel manifold, and Grassmann manifold [19]. The proposed MMFT minimized the marginal probability distribution by aligning different domains on the SPD manifold and learning Grassmann features in the Grassman manifold [18]. Three abovementioned manifolds can be
Proposed method
The MMFT framework is proposed for EEG signal classification in this section. The goal is to transfer the knowledge from multiple source subjects to the target subject. To takes advantage of the properties of Riemannian manifold and the knowledge transfer from multiple source domains, Section 3.1 introduces the MMFT framework. To solve the class imbalance problem, we proposed w-MMFT wherein the class with a smaller number of samples in the source domain gets a bigger weight in Section 3.2. To
Experiment and evaluations
In this section, we evaluate the performance of MMTF and LSA-MMFT on six EEG datasets as compared with several state-of-the-art algorithms. The code is available at https://github.com/rick604/MMFT. All evaluations are implemented in MATLAB R2017a on a laptop with i7-6700HQ CPU @ 2.60 GHz, 16 GB memory, and 64-bit Windows 10 Home Edition.
Conclusion
Transfer learning is a useful tool in EEG-based BCIs to tackle variations among different subjects and/or tasks. This paper considered offline unsupervised multi-source EEG classification. The proposed MMFT framework learned manifold features to minimize the marginal probability distribution, and then transferred manifold features using a voting mechanism. Moreover, we considered handling the class imbalance problem by proposing weighted MMFT (W-MMFT). The experimental results and its analogous
CRediT authorship contribution statement
Qingshan She: Methodology, Conceptualization, Funding acquisition. Yinhao Cai: Writing - original draft, Software, Validation. Shengzhi Du: Writing - review & editing, Formal analysis. Yun Chen: Project administration.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgements
This work was partly supported by Zhejiang Provincial Natural Science Foundation of China under Grant (No. LZ22F010003) and National Natural Science Foundation of China under Grant (Nos. 61871427, 61971168 and 62171171).
Qing-Shan She was born in Hubei Province, China. He received the B.E. degree in 2001 and M.E. degree in 2004, both from Lanzhou University of Technology, Lanzhou, China and the Ph.D. degree in 2007 from Zhejiang University, Hangzhou, China. He is currently a Professor with the School of Automation, Hangzhou Dianzi University, Hangzhou, China. His research interests include machine learning and pattern recognition, brain-computer interface and its applications.
References (47)
- et al.
Sub-band target alignment common spatial pattern in brain-computer interface
Comput. Methods Programs Biomed.
(2021) - et al.
Classification of covariance matrices using a riemannian-based kernel for bci applications
Neurocomputing
(2013) - et al.
Motor imagery eeg decoding using manifold embedded transfer learning
J. Neurosci. Methods
(2022) - et al.
The non-invasive berlin brain–computer interface: fast acquisition of effective performance in untrained subjects
NeuroImage
(2007) - et al.
Eeglab: an open source toolbox for analysis of single-trial eeg dynamics including independent component analysis
J. Neurosci. Methods
(2004) Brain–machine interfaces from motor to mood
Nature Neurosci.
(2019)- T.J.C. Limchesing, A.E.C. Chua, C.J.L. Shi, R.G. Baldovino, F.E.T. Munsayac, N.T. Bugtai, A review on recent...
Brain-computer interfacing: an introduction
(2013)- H. Cheng, Z. Wang, Z. Wei, L. Ma, X. Liu, On adaptive learning framework for deep weighted sparse autoencoder: A...
- et al.
Exponential synchronization for delayed dynamical networks via intermittent control: dealing with actuator saturations
IEEE Trans. Neural Networks Learn. Syst.
(2018)
Eegnet: a compact convolutional neural network for eeg-based brain–computer interfaces
J. Neural Eng.
A pso-based deep learning approach to classifying patients from emergency departments
Int. J. Mach. Learn. Cybern.
Transfer learning for eeg-based brain–computer interfaces: A review of progress made since 2016
IEEE Trans. Cognit. Develop. Syst.
Transfer learning: A riemannian geometry framework with applications to brain–computer interfaces
IEEE Trans. Biomed. Eng.
Multi-source based approach for visual domain adaptation
Experience selection using dynamics similarity for efficient multi-source transfer learning between robots
Combination of transferable classification with multisource domain adaptation based on evidential reasoning
IEEE Trans. Neural Networks Learn. Syst.
Domain adaptation via transfer component analysis
IEEE Trans. Neural Networks
Transfer feature learning with joint distribution adaptation
Mind the class weight bias: Weighted maximum mean discrepancy for unsupervised domain adaptation
Manifold embedded knowledge transfer for brain-computer interfaces
IEEE Trans. Neural Syst. Rehabil. Eng.
Cited by (18)
Manifold embedded instance selection to suppress negative transfer in motor imagery-based brain–computer interface
2024, Biomedical Signal Processing and ControlMulti-source adversarial transfer learning for ultrasound image segmentation with limited similarity
2023, Applied Soft ComputingSupervised feature selection on gene expression microarray datasets using manifold learning
2023, Chemometrics and Intelligent Laboratory SystemsCross-subject EEG emotion recognition using multi-source domain manifold feature selection
2023, Computers in Biology and Medicine
Qing-Shan She was born in Hubei Province, China. He received the B.E. degree in 2001 and M.E. degree in 2004, both from Lanzhou University of Technology, Lanzhou, China and the Ph.D. degree in 2007 from Zhejiang University, Hangzhou, China. He is currently a Professor with the School of Automation, Hangzhou Dianzi University, Hangzhou, China. His research interests include machine learning and pattern recognition, brain-computer interface and its applications.
Ying-Hao Cai was born in Zhejiang province, China. He received the B.E. degree from Jiangsu University, Zhenjiang, China, in 2020. He is currently pursuing the M.E. degree in control science and engineering with Hangzhou Dianzi University, Hangzhou, China. His research interests include transfer learning, brain-computer interface.
Sheng-Zhi Du received the M.S. degree in control theory and control engineering from Tianjin Poly Technology University, Tianjin, China, in 2001 and the Ph.D. degree in control theory and control engineering from Nankai University, Tianjin, China, in 2005. He is currently a professor in French South Africa Institute of Technology (F’SATI), Tshwane University of Technology, South Africa. His research interests include computer vision, AI, pattern recognition, and Human in the loop systems.
Yun Chen received the B.E. degree in 1999 from Central South University of Technology (Central South University), Changsha, China, and the M.E. degree in 2002 and Ph.D. degree in 2008, both from Zhejiang University, Hangzhou, China. He is currently a Professor with School of Automation, Hangzhou Dianzi University, Hangzhou, China. His research interests include stochastic systems, robust control and filtering, etc.
- ☆
Fully documented templates are available in the elsarticle package on CTAN.