Transfer learning in motor imagery brain computer interface: a review
image: Mingai Li, received her B.Sc. degree and M.Sc. degree from Daqing Petroleum Institute, Heilongjiang, China, in 1987 and 1990 respectively, and Ph.D. degree from Beijing University of Technology, Beijing, in 2006. She is currently a professor with the Faculty of Information Technology, Beijing University of Technology. Her current research interests mainly include brain computer interface, Artificial intelligent, pattern recognition and rehabilitation robot. Dongqin Xu, received her B.Sc. degree and M.Sc. degree from Hangzhou Dianzi University, Henan, China, in 2015 and 2018 respectively, and is currently pursuing her Ph.D. degree in Beijing University of Technology, Beijing, China. Her research interests include brain computer interface, data mining, knowledge transfer.
Credit: Mingai Li，Dongqin Xu.
“This review will lay a foundation for the popularization and in-depth research of transfer learning in BCI,” explained Dongqin Xu, the author of this research, “we believe that it will provide a quick and important reference for researchers engaged in transfer learning in BCI.”
Background
Stroke threatens the human health and life safety seriously; therefore, it has an important theoretical significance and application prospect to study intelligent and efficient motor nerve rehabilitation methods. Brain-computer interface technology based on motor imagery electroencephalography (MI-EEG) has demonstrated its specific advancement, effectiveness and brain-computer interaction in motor neurorehabilitation, and the accurate recognition of MI-EEG is the key. Recognizing MI-EEG by using deep learning technology can avoid or reduce the influence of feature engineering and human factors, but it faces the dilemma of small data amount and insufficient model training. However, each person has his/her own unique brain anatomy structure, function and neural activity pattern, which results in the obvious individual variability of MI-EEG, and also changes with MI tasks and time, thus the adaptability and generalization ability of deep learning methods are weakened.
Transfer learning, as a new machine learning methodology, may solve problems in related but different domains by using existing knowledge, and it is often applied to transfer training data from another domain for model training in case of insufficient training data. In recent years, an increasing number of researchers who engage in brain computer interface (BCI), have focused on using transfer learning to make most of the available electroencephalogram data from different subjects, effectively reducing the cost of expensive data acquisition and labeling as well as greatly improving the learning performance of the model. Combining with transfer learning makes it possible to borrow data or models.
Method
The classification accuracy of target domain is promoted by transferring the knowledge of source domain, and the main transfer learning methods include instance-based transfer learning, parameter-based transfer learning, and feature-based transfer learning. The instance-based transfer approach reuses some data of the source domain as a supplement to the target domain, that is, select the instances that are similar to the target domain from the source domain by using similarity measurement and add to the target domain. The parameter-based transfer approach uses the EEG signals of the source domain to pretrain a model, then directly transfer the model parameters or fine-tune the parameters with part of the target data to optimize the target model. The feature-based transfer learning approach refers to mapping instances from the source domain and target domain into a new data space, then extract features from this common space to minimize the domain divergence and improve the performance of the classification model on the target domain.
Difficulties
In the comparison experiments, different methods need to build different models, and the parameters of some methods are not very clear, which leads to differences in the reproduction of the results, if the code of the paper can be open accessed, it will bring convenience to the research work.
Future Work
In the future work, we will continue to study the decoding of motor imagery based on transfer learning, and combine functional near-infrared spectroscopy (fNIRS), and functional transcranial Doppler ultrasonography (fTCD) to conduct multimodal studies to improve the effectiveness of brain-computer interfaces in rehabilitation engineering.