Neural interfaces hold great potential to restore movement and communication function in millions of patients
with paralysis, neuromuscular disorders, traumatic brain injury, stroke, or communication disorders. These
systems rely on neural decoding algorithms to translate recorded neural activity into, for example, movements
of a prosthetic limb or intended speech sounds. However, many technical challenges limit the predictive
accuracy of these decoding models, preventing widespread deployment of restorative neuroprosthetic devices.
A key challenge is the limited data available to train decoding models. In existing approaches, models must be
trained separately for each individual subject, usually during simple behavioral tasks, and consequently fail to
generalize well to new subjects or complex naturalistic behavioral settings. This proposal leverages recent
advances in machine learning that directly address these limitations by developing a new decoding model
framework that is capable of combining neural data across many subjects and tasks, as well as incorporating
large-scale simulated data to improve prediction accuracy. In this framework, a single global model learns an
internal representation of the neural system that is invariant to variations in behavioral task and stimulus set,
anatomical variations, and functional variations in neural tuning of the underlying neuronal population (i.e.,
population-invariant neural decoding). This global model can therefore be applied and calibrated to new
subjects and behavioral tasks where little or no additional training data is available. Using an existing
intracranial neural data set collected from a large number of different subjects and stimulus sets, the project
will establish this new modeling approach based on deep learning architectures that are explicitly designed to
incorporate data pooled across subjects and tasks. We propose to show, through validation on measured
intracranial neural data, that a global neuronal population-invariant decoding model substantially improves
model prediction accuracy and generalization relative to existing state-of-the-art neural decoding models
across many subjects. Development and validation of this approach will open new avenues for researchers to
combine disparate data sets, for example, enabling community development, improvement, and sharing of
“open source” models that can be shared among research groups and effectively applied across research
studies. Thus, the proposal seeks to address key limitations in present neural decoding model approaches
which must ultimately convert measured neural activity into useful behavioral or communication parameters
across a wide range of subjects and complex behaviors in order for translational effects to be realized in
important clinical applications of neural interfaces.