Learning Interpretable Rules for Scalable Data Representation and Classification