Learning Transferable Representations For Visual Recognition