Brain-computer Interface Model Based on Convolutional Neural Networks and Adaptive Sample Reweighting

To solve the problem of information loss and overfitting caused by feature extraction, we establish a pure data-driven end-to-end convolutional neural network (CNN) model for electroencephalography (EEG) signals in the motion-imagination-type brain-computer interface. At the same time, to solve the problem that CNN requires a large amount of training data but a small amount of single-subject EEG data, we establish a method of using multi-subject data. By analyzing if other subject's data can improve the target subject's model, we eliminate those samples that contribute negatively to the target model. Then, in the training process of the CNN, we use a meta-learning technique to give each training data a weight. When training the CNN, after each step of the network parameter updating, we analyze the effect of the sample data in the training set on the final model and adaptively adjust the weight of each sample data. The experimental results show that our CNN can achieve better accuracy than traditional methods when using multi-person data and performing data cleaning and adaptive sample weighting techniques.