Hi, when I do my classification on a dataset with cross-entropy loss, I find the top1 accuracy about 91%, but top2 is as good as 96%,top3 improve a little，about 97%; Moreover, I find the big drop from top2 to top1 is due to several hard(easy to confusion) classes, AS shown in the following figures. My purpose only cares top1 accuracy, SO is there any way to reduce the gap between top2 and top1, thus to improve top1 accuracy.
