Regularised task‐related component analysis for robust SSVEP‐based brain‐computer interface

Abstract

Recent advancements in steady-state visual evoked potential-based brain-computer interface have been made possible by the widespread use of various spatial filtering methods. Task-related component analysis has superiority over existing subject-specific target frequency recognition methods. However, the optimised spatial filters of task-related component analysis generated from a small training dataset are susceptible to artefacts and noise, which can be overfitted, particularly in short time windows. To tackle this issue, the authors propose a regularised task-related component analysis that adopts three regularisation approaches to the objective function of task-related component analysis. Conventionally, the regularisation method is a simple and efficient way to overcome the overfitting problem, especially for a small training dataset. To this end, the proposed regularised task-related component analyses outperform the conventional task-related component analysis in terms of average classification accuracy and information transfer rate.