Research On MEG Decoding Based On Transfer Learning

Brain-computer interface is a direct communication and control channel that established between human brain and electronic devices without using the peripheral nerves and muscles.It makes it possible for humans to communicate with external environment through brain signals,allowing humans to express ideas or manipulate devices directly through the brain without the need for language or action.Magnetoencephalography is rapidly becoming an indispensable non-invasive brain imaging technology.Using professional instrumentation,magnetoencephalography can detect weak magnetic activity emanating from groups of neurons in the brain,and only magnetoencephalography can precisely localize and record these millisecond phenomena that produce signals approximately a billion times smaller than the Earth's magnetic field.Traditional magnetoencephalography decoding algorithms rely too much on the number of training samples,and the consistency of the distribution of training samples and test samples in the same feature space.In practical application,it is difficult to satisfy the above conditions,which limits the transferability of training data or training models between different subjects.In this study,the idea of transfer learning is applied to cross-subject magnetoencephalography decoding.By reviewing the transfer learning techniques that have achieved satisfactory results in brain decoding,three cross-subject magnetoencephalography decoding methods are proposed in this thesis.The specific research contents are as follows:In this study,based on the correspondence between points on Riemannian manifold and tangent vectors in tangent space,the same feature subspace between the sample covariance matrices of different subjects' magnetoencephalogram was found in tangent space to map,and the cross-subject magnetoencephalogram decoding based on Riemannian manifold learning was realized.In this study,each subject is regarded as a task,assuming that each subject's learning model has the same structure,by sharing the prior distribution information among model parameters,an improved multi-task learning framework is proposed based on Bayesian multi-task learning framework.On the basis of Riemannian manifold learning and the improved multi-task learning framework,a joint algorithm is proposed in this study.By combining the feature extraction of Riemannian manifold learning and the classification process of multi-task learning framework,the feature-model-based transfer is realized,and the performance of cross-subject magnetoencephalography decoding is further improved.In this experiment,magnetoencephalography datasets of 16 subjects in the task of target visual stimulus detection were used to verify the effectiveness of the above three algorithms.