Revisiting Unreasonable Effectiveness of Data in Deep Learning Era

   Abstract