Cross-subject federated transfer learning with quanvolutional layer for Motor Imagery classification

• Published in: Chinese Automation Congress (Online) pp. 5736 - 5741
• Main Authors: Hu, Ruihan, Zhou, Xuefeng, Xu, Zhihao, Liao, Zhaoyang, Wu, Hongmin, Qu, Hongyi, Tang, Zhi-Ri
• Format: Conference Proceeding
• Language: English
• Published: IEEE 22.10.2021
• Subjects: Cross-subject transfer learning; Encoding; Feature extraction; FederatedAveraging; Manifolds; Medical treatment; Quantum computing; The Riemannian spatial Encoder-Decoder backbone; Transfer learning; Transportation
• ISSN: 2688-0938
• DOI: 10.1109/CAC53003.2021.9727351

The Brain-Computer Interface (BCI) systems play an important role in the Rehabilitation therapy, Smart-home, Intelligent Transportation fields. To the best of our knowledge, different from the cross-trial and cross-run tasks, the data privacy of the huge amount of datasets from multiple subjects prevents the research for the cross-subject transfer learning of the BCI classification task. In this paper, a simple federated transfer framework, namely Federated Transfer Network with Quanvolutional Architecture (FTL-QL), is proposed to overcome this problem. The Riemannian spatial Encoder-Decoder backbone that contains the Quanvolutional and Encoder-Decoder layers to execute the quantum, Manifold Riemannian Coding and Log-Euclidean Riemannian Decoding computation to extract the discriminative information features for cross-subjects' transfer learning. Then the Federated module which calculated by the FederatedAveraging method to train the top layer of the FTL-QL for each subject. The performance of the FTL-QL is benchmarked on the EEG Motor Imagery datasets. Several experiments about the BCI classification task show the proposed FTL-QL can achieve superior learning performance for Cross-subject transfer learning.