Research On Algorithms Based On Variational Information Encoding And Convolutional Neural Networks For Signal Classification In Brain-Computer Interface

People can interact with devices without any muscle activity with brain-computer interface,which is one of the important research fields of brain-computer intelligence and brain-like computing.The analysis and processing of electroencephalogram signals is key research topic for the practical implementation of brain-computer interface.However,high data dimension,susceptible to noise and low signal-to-noise ratio in electroencephalogram signals have brought serious challenges to signal detection and recognition.And traditional detection models deal with noise in the data preprocessing process,which rarely reduce the impact of noise on recognition with problem scene and cannot effectively extract features.In this paper,we built a series of effective classification models based on analysis with deep neural networks and information theory.Firstly,an event-related potential detection algorithm based on variational auto-encoder and convolutional neural network is proposed.There are two steps in the algorithm: characterization and classification.The characterization process is completed by a variational auto-encoder with the structure of convolutional neural networks.The characterization network map the information of electroencephalogram signals into hidden variables with reconstruction task.And the hidden variables are classified by the classification network.According to experimental results and discussions,this model is an effective electroencephalogram signal classification model.Secondly,an event-related potential detection algorithm based on variational information bottleneck and convolutional neural networks is proposed.The electroencephalogram data is encoded by the variational information bottleneck based on the convolutional neural networks.And then the hidden variables are classified.These two progress are optimized together.The model can adjust the information transmission in the network with the network structure and loss function.The variational information bottleneck retain the information related to the target stimulus in electroencephalogram signals as much as possible while limiting the transfer of the information that is not helpful for classification in the network.The experimental results show that the algorithm has excellent classification ability,and surpasses lots of algorithms when the data is enough.The mechanism of variational information bottleneck int the model is elaborated with experimental results.Moreover,the two models are unified in framework of variational information coding.The differences of mechanism and performance between the two models are analyzed with theoretical and experimental aspects.These two models are truly practical examples of the combination of information theory,probability theory,Bayesian theory and deep neural networks,which provides new ideas for the study of the interpretability of deep neural network and solutions to classification problems for electroencephalogram signals.Our work is promising for the brain science and neuroscience.