Deep James-Stein Neural Networks for Brain-Computer Interfaces
Nonparametric regression has proven to be successful in extracting features from limited data in neurological applications. However, due to data scarcity, most brain-computer interfaces still rely on linear classifiers. This work leverages the robustness of the James-Stein theorem in nonparametric regression to harness the potentials of deep learning and foster its successful application in neural engineering with small data sets. We propose a novel method that combines James-Stein regression for feature extraction, and deep neural network for decoding; we refer to the architecture as deep James-Stein neural network (DJSNN). We apply the DJSNN to decode eye movement goals in a memory-guided visual saccades to one of eight target locations. The results demonstrate that the DJSNN outperforms existing methods by a substantial margin, especially at deep cortical sites.
