Tensor Decompositions of EEG Signals for Transfer Learning Applications

Sammanfattning: In this report, tensor decomposition methods of EEG signals have been evaluated for the purpose of transfer learning. The aim has been to address the person-to-person Brain-Computer Interface (BCI) calibration problem by transferring training data between sessions, which can shorten calibration times, extend the amount of training data, and enable using data from simulated environments in real world applications. For this, the datasets AlexMI (binary motor imagery) and SA Driving (drowsiness detection during simulated driving) have been analyzed. Tensor decompositions were performed unsupervised in two pipelines, with aim of capturing universal structures relevant to BCI tasks. For the first pipeline, two decompositions (Canonical Polyadic and Tucker) were computed to compare similarity between sessions. From that, a subset of sessions were selected that during classification, were aimed to outperformrandom selection and training with the full training database. In the first pipeline, a new similarity measure was designed, which included weighting of the factor matrix in the mode of interest. We consider this a more representative measure of how similar two sessions are, compared to simply studying the unweighted factor matrices, which was done in previous literature. For the second pipeline, one tensor decomposition (Tucker) was used for feature extraction and similarity comparison between sessions. The aim was the same as for pipeline one, with the addition of investigating the properties of tensor decompositions as features. The results show that unsupervised tensor decompositions can extract structures of varying relevance to a classification problem but did not result in superior performance when used as features.With this knowledge, we propose extending tensor decompositions to supervised and/or nonlinear ones. Additionally, the proposed session selection methods showed potential in classification, but no significant conclusions could be drawn of their superiority compared to random selection or trainingwith the full database. Additionally, the classifiers had a large variation in performance between sessions, making them far from applicable for a BCI in a real-world environment today.