Model Drift in Brain-Computer Interfaces and Space-Based Remote SensingdissertationsKhoshnevis, MonacreatorSandstede, BjornReaderSimeral, JohnReaderKellner, JamesReaderBrown University. Department of Applied Mathematicssponsor2024xv, 170 p.born digitalThesis (Ph. D.)--Brown University, 2024thesesAs predictive machine learning algorithms become increasingly prevalent, addressing performance degradation due to model drift throughout a system's life cycle has become critical. This thesis explores model drift detection, understanding, and adaptation across various applications. First, we introduce two hypothesis tests designed to detect statistically significant drift in probabilistic encoding models applied in Brain-Computer Interfaces (BCIs). Our proposed p-values specifically target model drift in the conditional probability distribution of neural signals given the motor imagery task and are robust to changes in the neural marginal distribution. Then we study this marginal distribution shift between neural data by leveraging a distribution-based drift detection method in BCIs that employs an information-theoretic distance, and we investigate the relationship between this measure and decoder performance. Third, to adapt to a specific type of model drift in multiple-task BCIs, such as point-and-click, we propose two statistical approaches: a data-dependent regularization technique and a method that identifies a neural projection space where the primary task linear decoder remains robust to any anticipated model drift. Finally, we present two methods to adapt to model drift related to NASA’s Global Ecosystem Dynamics Investigation (GEDI) mission, which arises from measurement error where only a surrogate measurement of input data is available. Our approach addresses the challenge of transferring from surrogate inputs to real features.Brain-computer interfacesGEDIDrift--Mathematical modelsEnglishRPB20241015