BCIs for Everyone, Everyday: Generalized Machine Learning Models for Decoding of Human Brain Data

Abstract

Brain Computer Interfaces (BCIs) offer immense potential to enhance the quality of life for individuals worldwide, spanning applications in prosthetics, neurofeedback, and virtual reality. Recent advancements in deep learning techniques and the exponential growth of available data have significantly improved BCI research and neural decoding over the last decade. However, there remains a critical need to extend the applicability of these systems to everyone, everyday. While much of the current BCI research has focused on developing neural decoders for individuals performing specific experimental tasks, the true potential of BCIs lies in their real-world applicability. To achieve broad BCI use, we must enhance key aspects of BCI systems, such as their ability to generalize across users, be robust to natural behaviors, and provide inclusive user experiences. This thesis addresses these challenges by exploring various components of BCI systems to improve their real-world applicability. Chapter 4 introduces HTNet, a deep learning neural decoder for ECoG data that is robust to naturalistic movement behaviors in humans, and individual variability. Building on this work, Chapter 5 extends the HTNet model to cognitive behaviors in EEG, highlighting its task-relevant robustness. Chapter 6 explores the structure of naturalistic human neural data by extracting estimates of neural manifolds and describing their geometry. We find that naturalistic neural activity resides within similarly aligned low-dimensional neural manifolds across movement types, days and individuals, potentially indicating a shared manifold for natural behavior. Chapter 7 explores various strategies to enhance the robustness of neural decoders to naturalistic human neural data. These strategies include decoding complex movement information, leveraging unlabeled naturalistic neural data through self-supervised learning, and decoding from neural manifolds. While the majority of this thesis focuses on machine learning improvements for BCI systems, Chapter 8 explores another aspect of BCI systems – the user experience. This chapter investigates gender gaps in BCI performance and uses human-computer interaction techniques to understand how aspects of the user interface influence BCI performance. In the concluding chapter, we propose suggestions for future research to further develop innovative neural decoders that are not only robust to naturalistic behaviors and individual variability, but also designed with everyone in mind.