Finger Movements Classification using Autonomous Transfer Learning

Abstract

Misclassification resulting from data shifts is one of the difficulties in classification systems based on brain-computer interfaces. This may occur if the classification system does not consider distributional shifts between training and test sets of data. The performance of the qualification system in the EEG classification domain may significantly deteriorate because of inter-session trials. This article introduces autonomous transfer learning (ATL) in the pipeline for classifying finger movement based on EEG motor imagery. On same-session trials, the average accuracy of four subjects is 0.504, but on inter-session trials, it is slightly worse (0.475). Due to its simplicity, the suggested method also has the fastest processing speed and offers the opportunity to be used on edge devices where iterative training is not feasible.