On Tackling Fundamental Constraints in Brain-Computer Interface Decoding via Deep Neural Networks

Abstract

A Brain-Computer Interface (BCI) is a system that provides a communication and control medium between human cortical signals and external devices, with the primary aim to assist or to be used by patients who suffer from a neuromuscular disease. Despite significant recent progress in the area of BCI, there are numerous shortcomings associated with decoding Electroencephalography-based BCI signals in real-world environments. These include, but are not limited to, the cumbersome nature of the equipment, complications in collecting large quantities of real-world data, the rigid experimentation protocol and the challenges of accurate signal decoding, especially in making a system work in real-time. Hence, the core purpose of this work is to investigate improving the applicability and usability of BCI systems, whilst preserving signal decoding accuracy.

Recent advances in Deep Neural Networks (DNN) provide the possibility for signal processing to automatically learn the best representation of a signal, contributing to improved performance even with a noisy input signal. Subsequently, this thesis focuses on the use of novel DNN-based approaches for tackling some of the key underlying constraints within the area of BCI. For example, recent technological improvements in acquisition hardware have made it possible to eliminate the pre-existing rigid experimentation procedure, albeit resulting in noisier signal capture. However, through the use of a DNN-based model, it is possible to preserve the accuracy of the predictions from the decoded signals. Moreover, this research demonstrates that by leveraging DNN-based image and signal understanding, it is feasible to facilitate real-time BCI applications in a natural environment. Additionally, the capability of DNN to generate realistic synthetic data is shown to be a potential solution in reducing the requirement for costly data collection. Work is also performed in addressing the well-known issues regarding subject bias in BCI models by generating data with reduced subject-specific features.

The overall contribution of this thesis is to address the key fundamental limitations of BCI systems. This includes the unyielding traditional experimentation procedure, the mandatory extended calibration stage and sustaining accurate signal decoding in real-time. These limitations lead to a fragile BCI system that is demanding to use and only suited for deployment in a controlled laboratory. Overall contributions of this research aim to improve the robustness of BCI systems and enable new applications for use in the real-world.