Deciphering Raw Data in Neuro-Symbolic Learning with Provable Guarantees