Study On Bayesian-uncertainty-estimation Based Deep Learning For Brain-computer Interface

Brain computer interface(BCI)allows users to communicate and interact with external devices through brain signals,providing a new enhanced communication technology for people who are paralyzed or have other severe motor disabilities.Among them,the P300-based character speller is one of the most widely studied BCI applications.To improve the performance of such BCI systems,deep learning technology has been widely used in P300 detection and achieved remarkable results.However,we can only obtain point estimates of the weights in traditional neural networks.These networks cannot effectively measure the prediction uncertainty,so we cannot know whether the prediction is reliable or not.Recent studies have shown that the Bayesian neural network(BNN)can capture the model uncertainty by putting the probability distribution over the model weights,so it can provide effective prediction uncertainty.On the other hand,BNN introduces prior information to the model weights,so it is robust to overfitting.In view of these advantages,this paper studies how to use BNN to solve the uncertainty estimation problem of deep learning on P300 brain-computer interface,and explores its applications in P300 detection,adversarial attack defense,optimization of deep network structure,rejection classification problem,and deep learning on small datasets.The research content of this paper is mainly divided into the following three aspects:(1)A P300 detection method based on Bayesian convolutional neural network is proposed.The weights in the convolutional neural network are represented by probability distributions.In the prediction stage,a certain number of sub-networks are sampled by Monte Carlo sampling,and the predictions of these sub-networks are integrated.This not only makes the prediction more reliable and improves the performance of P300 detection,but also obtains the prediction uncertainty.At the same time,weight uncertainty is used to study the pruning of deep networks.In addition,prediction uncertainty is used to further study the defense performance against adversarial attacks,so as to improve the security of character spellers.(2)The traditional capsule network cannot measure the prediction uncertainty,and it has a large number of parameters,so it is easy to overfit on EEG datasets.To solve these problems,this paper proposes a Bayesian capsule network for P300 detection.The weights can be regularized by putting the prior distribution over the model weights,thereby reducing the risk of overfitting on small EEG datasets.The robustness of this network is investigated in adversarial attack experiments.In addition,this paper designs a classification rejection strategy,which can reduce the classification error rate of the model by rejecting the predictions with high uncertainty.The performance of the network under small EEG datasets is verified by reducing the training samples.(3)The depth of the network plays a key role in P300 feature extraction.How to determine the optimal depth of the network has always been a difficult problem in P300 deep learning.To solve this problem,a temporal convolution depth uncertainty network is proposed to perform probabilistic reasoning for the depth of temporal convolution in P300 detection.After training,a posterior probability can be obtained on each temporal convolutional layer,which reflects the importance of each temporal convolutional layer for P300 detection.These probabilities can be used to optimize the depth of the network,thus improving the efficiency of the character speller.Experiments show that the network can achieve excellent P300 detection performance.More importantly,the posterior probability can be used to guide the design of the number of network layers in BCI deep learning.