A New Architecture for Deriving Dynamic Brain-Machine Interfaces

Great potential exists for future Brain Machine Interfaces (BMIs) to help paralyzed patients, and others with motor disabilities, regain (artificial) motor control and autonomy. This paper describes a novel approach towards the development of new design architectures and research test-beds for advanced BMIs. It addresses a critical design challenge in deriving the functional mapping between the subject’s movement intent and actuated behavior. Currently, adaptive signal processing techniques are used to correlate neuronal modulation with known movements generated by the subject. However, with patients who are paralyzed, access to the individual’s movement is unavailable. Inspired by motor control research, this paper considers a predictive framework for BMI using multiple adaptive models trained with supervised or reinforcement learning in a closed-loop architecture that requires real-time feedback. Here, movement trajectories can be inferred and incrementally updated using instantaneous knowledge of the movement target and the individual’s current neuronal activation. In this framework, BMIs require a computing infrastructure capable of selectively executing multiple models on the basis of signals received by and/or provided to the brain in real time. Middleware currently under investigation to provide this data-driven dynamic capability is discussed.