An intracortical brain-computer interface (iBCI) is used to record electrical signals directly from a person's
brain, predict their intention from those signals, then control an assistive device (e.g., a computer cursor,
prosthetic limb, or powered wheelchair) according to those intentions. This technology enables severely
paralyzed people to interact with the world. However, designing robust algorithms to extract intent from
recordings of single neurons is extremely challenging, in large part because of the very limited access to
humans, or even monkeys, from whom these invasive recordings can be made.
In this project, we will develop a model iBCI system that generates real-time biomimetic neural data by
capturing the high-degree-of-freedom finger movements of able-bodied human subjects. To accomplish this,
we will construct a modular recurrent neural network (RNN). The RNN will be trained to predict the motor
cortex activity of a monkey from the monkey's own finger kinematics. Small modules of the RNN will be
interchanged according the particular animal or recording session to model the high inter-session variability
present in motor cortex. Once the modular RNN is trained, its weights will be fixed and human finger
kinematics will be used as the RNN inputs, which will generate subject-controlled emulated neural activity. The
emulated neural activity can be passed to iBCI decoding algorithms that control computer cursors or other
physical devices, allowing human subjects to interact directly with decoders in real time, closed-loop
conditions. We call this model system the jaBCI. The jaBCI is low cost and noninvasive, making it possible to
rapidly test and design novel iBCI decoders using statistically rigorous sample sizes.
The project will be executed in close collaboration with intracortical microelectrode array data expert Dr. Lee
Miller at Northwestern University. Dr. Miller's lab, with the help of our consultant Dr. Mathis, will obtain
simultaneous finger kinematics and neural activity of monkey subjects that will serve as the training data for the
RNN component of the iBCI model.
We will validate the emulated neural data generated by the jaBCI across many measures to ensure the model
captures as many features of intracortical data as possible. These include comparing the model and actual
iBCI in subject performance, learning rates, control strategies, neural variation across days, neural firing rate
distributions, and low-dimensional neural dynamics. With the validated model, we will undertake a study to
rigorously evaluate the highest performing, current state-of-the-art iBCI decoders. This will yield useful insight
into the features of decoders that yield the greatest performance gains, overcoming the current impossibility to
compare iBCI decoders in well-controlled studies using more than two or three naïve human subjects. We will
also use the iBCI model to evaluate novel decoder designs, and to determine the features of neural dynamics
that are consistent across common iBCI tasks to help focus decoder development on those features.