A machine learning approach to taking EEG-based brain-computer interfaces out of the lab

Abstract

Despite being a subject of study for almost three decades, non-invasive brain- computer interfaces (BCIs) are still trapped in the laboratory. In order to move into more common use, it is necessary to have systems that can be reliably used over time with a minimum of retraining. My research focuses on machine learning methods to minimize necessary retraining, as well as a data science approach to validate processing pipelines more robustly. Via a probabilistic transfer learning method that scales well to large amounts of data in high dimensions it is possible to reduce the amount of calibration data needed for optimal performance. However, a good model still requires reliable features that are resistant to recording artifacts. To this end we have also investigated a novel feature of the electroencephalogram which is predictive of multiple types of brain-related activity. As cognitive neuroscience literature suggests, shifts in the peak frequency of a neural oscillation – hereafter referred to as frequency modulation – can be predictive of activity in standard BCI tasks, which we validate for the first time in multiple paradigms. Finally, in order to test the robustness of our techniques, we have built a codebase for reliable comparison of pipelines across over fifteen open access EEG datasets.