Rethink Deep Learning with Invariance in Data Representation