最大熵模型中的heldout accuracy（张乐老师提供的最大熵模型包相关技术）
里面有一段这样的描述，我想了解一下有关heldout的相关详情，并且为什么会因此引入高斯方差？谢谢！
In this example, it seems performance peaks at iteration 9. Further training actually brings down the accuracy on
the heldout data, although the training accuracy continues to increase. Applying a Gaussian prior can help avoid
overfitting, just use -g float to specify the global Gaussian variance σ 2 .
