Recognition of MI-EEG Signals by Extended-LSR-based Inductive Transfer Learning

Zhibin Jiang ^1,2 Keli Hu ^1,3 Jia Qu ⁴ Zekang Bian ^5,6 Donghua Yu ^1,2 Jie Zhou ^1,2*

• ¹ Department of Computer Science and Engineering, Shaoxing University, Shaoxing, Zhejiang Province, China
• ² Institute of Artificial Intelligence, Shaoxing University, Shaoxing, China
• ³ Information Technology R&D Innovation Center of Peking University, Shaoxing, China
• ⁴ School of Computer Science and Artificial Intelligence, Changzhou University, Changzhou, China
• ⁵ Department of AI & Computer Science, Jiangnan University, Wuxi, China, Wuxi, China
• ⁶ Department of Taihu Jiangsu Key Construction Lab. of IoT Application Technologies, Wuxi, China

Motor imagery electroencephalographic (MI-EEG) signal recognition is used in many brain-computer interface (BCI) systems. In most existing BCI systems, this identification relies on a classification algorithm. However, generally, a large amount of subject-specific labeled training data is needed to reliably calibrate the classification algorithm for each new subject. To over-come this challenge, an effective strategy is to introduce transfer learning into the construction of intelligent models, in which knowledge is learned from the source domain to enhance the performance of the model trained in the target domain. Although transfer learning has been used in EEG signal recognition, many existing transfer learning methods are designed only for a specific intelligent model, which makes these methods lack application and generalization. To extend the scope of application and generalization, an extended-LSR-based inductive transfer learning method is proposed to realize transfer learning for several classical intelligent models, including neural networks, Takagi-Sugeno-Kang(TSK) fuzzy systems, and kernel methods. The proposed method not only can be used to transfer useful knowledge from the source domain to boost learning performance in the target domain when the training data of the target domain are insufficient but also has better application and generalization with integrating multiple classic base models. The experimental results demonstrate the effectiveness of the proposed method in MI-EEG signal recognition.