It was found that with the increase of S1, the correct rate would decrease oppositely.

• 我们发现当S1较大时，识别正确率反而下降，可能此时网络的非线性程度过高，使其不适合于该类样本集的训练。