UC San Diego

Neural interfaces enable interaction with the brain through direct recordings of neural activity. For individuals suffering from conditions limiting their ability to interact with and experience the world, solutions leveraging neural interfaces restore function and improve quality of life. Recent advances in brain-machine interfaces have enabled applications of such technology to achieve functional restoration in subjects with conditions spanning anarthria and tetraplegia. While these solutions begin to restore lost function, neural decoders require further improvement to match fully-able human performance.

Building robust neural decoders generally involves leveraging volitional responses, commonly motor intent, to train and deploy a model. Importantly, once a model is deployed, it is only updated in limited contexts (e.g. calibration). This can lead to loss of utility over time due to neural, measurement, and environmental sources. Using augmentative non-volitional responses, meaningful context can be provided to controllers online for adaptation and retraining to improve decoder robustness and usability.

In this work, we explore the full neural decoder development process. We first address challenges spanning clinical and laboratory data collection using a novel flexible experiment framework. We then critically review deep feature transforms and their utility in neural decoders. We next extend LFADS to intracranial human recordings and evaluate its application learning neural dynamics of reward. Finally, we seek to improve the algorithmic exploration and development process by demonstrating a pathway for scaling novel algorithms and improving disseminability. Through this work, we work towards further enabling high-performance neural prostheses.