Distilling Symbolic Priors for Concept Learning into Neural Networks