面向运动想象分类任务的任务导向子域对抗迁移网络

Task-oriented Subdomain Adversarial Transfer Networks for Motor Imagery Classification Tasks

运动想象是一种应用前景广泛的脑机接口范式.在基于脑电的运动想象分类任务中,由于设备和被试的缘故,会导致与被试、时间相关的数据分布漂移现象.这种数据分布漂移会使得分类器分类精度下降.而迁移学习能很好地解决这种分布漂移现象.本文提出了一种新的单源域选择算法,多子域可迁移性估计(multi-subdomain transferability estimation, MSTE)和一种新的迁移方法,任务导向的子域对抗迁移网络(task-oriented subdomain adversarial transfer network, ToSAN),用于脑电信号的分类任务. MSTE能评估源域和目标域在时间和类别上的相似性. ToSAN能面向分类任务分解特征,在与任务相关的特征上进行多个子域对齐,从而克服分布差异.在BCI Competition IV 2a和BCI Competition IV 2b上的实验结果表明, ToSAN相比于其他方法在分类准确率上提高了最少2.67%, 8.6%. MSTE和ToSAN的结合在BCI Competition IV 2a和BCI Competition IV 2b数据集上分别达到了81.73%和88.73%的分类准确率,显著优于所有对比方法.

Motor imagery is a promising brain-computer interface paradigm.In the motor imagery classification tasks based on EEG,the equipment and the subjects will lead to the phenomenon of data distribution drift related to the subjects and time.This data distribution drift will reduce the classification accuracy of the classifier.Transfer learning can solve this distribution drift phenomenon very well.In this study,a new single source domain selection algorithm,multisubdomain transferability estimation(MSTE)and a new transfer method,task-oriented subdomain adversarial transfer network(ToSAN),for the classification tasks of EEG signals are proposed.MSTE can evaluate the similarity in time and category between the source domain and the target domain.ToSAN can decompose features for classification tasks and perform multiple subdomain alignments on task-related features to overcome distribution differences.The experimental results on BCI Competition IV 2a and BCI Competition IV 2b show that compared with other methods,ToSAN improves the classification accuracy by at least 2.67%and 8.6%,respectively.The combination of MSTE and ToSAN achieve a classification accuracy of 81.73%and 88.73%on the BCI Competition IV 2a and BCI Competition IV 2b datasets,which is significantly better than all comparison methods.