A polar prediction model for learning to represent visual transformations