Semi-supervised Classification Of Brain Signals Based On VAE Framework

Recent years,Brain-Computer Interface(BCI)technology has developed rapidly.It can realize the user's brain to directly control the external device through human-computer interaction.The core part of its technology lies in the recognition of brain signal patterns.However,when the brain-computer interface implements pattern classification,a large amount of data is often required to train the classifier.But compared with the acquisition of brain signal data,it is time-consuming and laborious to assign the label information to these data.Therefore,the introduction of a semi-supervised learning algorithm,using the combination of labeled and unlabeled data,can effectively reduce the training cost of the braincomputer interface system.In this thesis,a semi-supervised learning algorithm for P300 signal classifier training is designed on the basis of a probability graph model,and the structure is specifically improved to improve the model interpretability of the semi-supervised algorithm in this application direction.The main research contents of this article are as follows:(1)This thesis designs a P300 signal semi-supervised learning algorithm based on Variational Autoencoder(VAE).Correlate the input signal,label and hidden space by means of probability graph,effectively use both labeled and unlabeled data.The neural network-based model can reduce the process of feature extraction in traditional semi-supervised methods.In addition,this thesis analyzes the problems caused by the imbalance of positive and negative samples in the P300 EEG data set,and designs a new loss function to solve it.(2)In this thesis,the concept of energy-based models(Energy-Based Models,EBMs),combined with the loss function of the above-mentioned VAE-based semi-supervised algorithm,is to design the energy-based regularization,the purpose is to limit the degree of freedom of the model Improve the accuracy of semi-supervised learning in the case of small samples.The experimental results show that the semi-supervised learning algorithm based on VAE proposed in this thesis can effectively implement the semi-supervised learning of P300 signal classifier.After adding energy-based regularization,the accuracy of the model in small samples has been greatly improved,and the convergence time of model training has been significantly shortened.At the same time,this thesis compares and analyzes the spatial features and the P300 feature signals.Compared with traditional supervised or semi-supervised learning methods,the method proposed in this thesis has significant advantages in model interpretability.