Improving BCI generalizability with multi-task modeling and autocalibration

Project Abstract

Brain-computer interfaces (BCIs) are systems that enable using neural activity to control and interact with external devices. For people who lose the ability to move or speak due to injury or disease, BCIs provide a potential avenue to restore this loss of function. Recent examples are showing the remarkable extent of this potential, in which people with paralysis have been able to use BCIs in complex settings such as text typing or speech generation. Here we propose work to improve the ability of BCIs to generalize to new tasks and adapt to changes in the underlying neural signal over time. Success in these two areas will enable BCIs to be used in widespread settings for long durations. First, we propose to develop BCI systems that are trained to accomplish a variety of tasks. We expect that such a system will learn a general representation of the neural signal that will enable quick adaption to new tasks with little supervision. Next, we also propose methods to automatically calibrate BCIs over time. Due to a variety of factors, the measured neural activity by the BCI changes over time. BCIs must be calibrated to adapt to these changes to achieve strong performance. Here we will work towards calibration methods that do not rely on new supervised training data from a new day or session of use. As part of this work, we’ll develop statistical models of shared structure and variability in neural activity over time. Together, our proposed research directions will help support the development of long-term, general purpose BCIs for clinical use.