Extensions and Application of the Robust Shared Response Model to Electroencephalography Data for Enhancing Brain-Computer Interface Systems

摘要

Brain Computer Interfaces (BCI) decode electroencephalography (EEG) data collected from the human brain to predict subsequent behavior. While this technology has promising applications, successfully implementing a model is challenging. The typical BCI control application requires many hours of training data from each individual to make predictions of intended activity specific to that individual. Moreover, there are individual differences in the organization of brain activity and low signal-to-noise ratios in noninvasive measurement techniques such as EEG. There is a fundamental bias-variance trade-off between developing a single model for all human brains vs. an individual model for each specific human brain. The Robust Shared Response Model (RSRM) attempts to resolve this tradeoff by leveraging both the homogeneity and heterogeneity of brain signals across people. RSRM extracts components that are common and shared across individual brains, while simultaneously learning unique representations between individual brains. By learning a latent shared space in conjunction with subject-specific representations, RSRM tends to result in better predictive performance on functional magnetic resonance imaging (fMRI) data relative to other common dimension reduction techniques. To our knowledge, we are the first research team attempting to expand the domain of RSRM by applying this technique to controlled experimental EEG data in a BCI setting. Using the openly available Motor Movement/ Imagery dataset, the decoding accuracy of RSRM exceeded models whose input was reduced by Principal Component Analysis (PCA), Independent Component Analysis (ICA), and subject-specific PCA. The results of our experiments suggest that RSRM can recover distributed latent brain signals and improve decoding accuracy of BCI tasks when dimension reduction is implemented as a feature engineering step. Future directions of this work include augmenting state-of-the art BCI with efficient reduced representations extracted by RSRM. This could enhance the utility of BCI technology in the real world. Furthermore, RSRM could have wide-ranging applications across other machine-learning applications that require classification of naturalistic data using reduced representations.