This paper was accepted on the workshop “Machine Studying 4 Bodily Sciences” at NeurIPS 2022.
Hybrid modelling reduces the misspecification of knowledgeable bodily fashions with a machine studying (ML) part discovered from information. Equally to many ML algorithms, hybrid mannequin efficiency ensures are restricted to the coaching distribution. To handle this limitation, right here we introduce a hybrid information augmentation technique, termed knowledgeable augmentation. Primarily based on a probabilistic formalization of hybrid modelling, we exhibit that knowledgeable augmentation improves generalization. We validate the sensible advantages of knowledgeable augmentation on a set of simulated and real-world programs described by classical mechanics.
