Contrastive Learning Neuromotor Interface From Teacher

Adaptive neuromotor interfaces are poised to enhance human-computer interaction experience and increase human mobility and accessibility to diverse users. These interfaces are important in assisting human users with dynamic control tasks where direct, physical interactions may be undesirable or impossible by predicting user intent from neuro signals. Existing methods for developing neuromotor interfaces suffer from distribution shifts due to inter and intra-user variability and the requirement for large amounts of supervised training data. Towards implicitly adapting to streaming user behavior \emph{without} intent labels, we propose an interactive contrastive fine-tuning method to address these limitations. We formulate pseudo-intent labeling as a Bayesian inference problem guided by an optimal task policy referred to as the "teacher" prior. Using a simulated robotic control task, we show that our method successfully aligns with user intent even when the teacher prior is misspecified against a diverse group of simulated users.