Harnessing Function Space of Machine Learning Models for Interpretability, Generalizability, and Efficiency