Study On The Classification Of EEG Based On Multi-task Learning

In the field of artificial intelligence,Brain-Computer Interface(BCI)technology has gradually become a topic of great concern.The interaction with the external environment can be accomplished by transforming the EEG signals collected from the scalp into control commands that directly control the computer or external devices without relying on the human peripheral nervous system,language,and limb movements.This new type of human-computer interaction technology can provide a way for patients with normal thinking but serious movement disorders to communicate and control with the outside world.It has great application value in the field of medical rehabilitation,and also has potential in new entertainment,industrial,military and other fields.Practical value.The EEG signals processed by the brain-computer interface system are complex signals with nonlinear non-stationary characteristics.The statistical distribution of EEG signal data collected by the same subject at different times may have large differences.The assumption that the traditional machine learning model has independent and identical distribution between training samples and test samples is often not established in EEG signal processing,which limits the reusability of EEG signal training data.Usually,in order to train a stable and reliable classification model,the brain-computer interface system needs to carry out a long-time training sample collection process in the experiment,which is time consuming and boring,and imposes a great burden on the subject.How to reduce the number of training samples,shorten the experiment time,reduce the experimental burden of the subject and ensure the performance of the classification model is one of the difficulties faced by the brain-computer interface technology.This paper proposes a multi-task learning feature extraction algorithm and a multi-task principal component analysis algorithm around the above problems faced by the braincomputer interface system.The main research results are as follows:(1)Propose a multi-task principal component analysis algorithm.Based on the sample difference matrix of the training set and test set of multiple tasks,this paper extracts the feature subspace shared between the sample difference matrices,and then uses this subspace as the common knowledge structure between tasks in multiple principal component analysis tasks.The transfer is performed such that in each principal component analysis task,after the training data and the test data are reduced to a low-latency subspace,the distribution difference between the training set and the test set sample is small in the low-latitude subspace.Reduce the data dimension to solve the problem of too few training samples.(2)Propose a multi-task common spatial pattern algorithm.This algorithm extracts the subspaces with good classification ability shared by multiple tasks,and the orthogonal complement space of the subspace is used as the penalty function in the form of penalty function in all common spatial model algorithm tasks.Passing,sharing the classification information between the more complete retention of multiple tasks while making full use of multiple task data to learn a classifier with better generalization performance;The experimental results show that these two methods achieve a reliable classification model with a small number of training samples,and finally achieve better classification accuracy than single-task training.