深度神经网络在测试时面对如此大的网络是很难克服过拟合问题的。
Dropout能够很好地解决这个问题。通过阻止特征检测器的共同作用来提高神经网络的性能。这种方法的关键步骤在于训练时随机丢失网络的节点单元包括与之连接的网络权值。在训练的时候，Dropout方法可以使得网络变得更为简单紧凑。在测试阶段，通过Dropout训练得到的网络能够更准确地预测网络的输出。这种方式有效的减少了网络的过拟合问题，并且比其他正则化的方法有了更明显的提升。
本文通过一个简单的实验来比较使用Dropout方法前后网络的性能优劣情况。(It is difficult to overcome the problem of fitting a deep neural network in the face of such a large network in testing.
Dropout can solve this problem well. The performance of the neural network is improved by preventing the common action of the feature detector. The key step of this method is that the node unit of the random loss network consists of the network weights connected to it during the training. In training, the Dropout method can make the network more compact. In the test phase, the network trained by Dropout can predict the output of the network more accurately. This method effectively reduces the over fitting problem of the network, and has a more obvious improvement than the other regularization methods.
In this paper, a simple experiment is used to compare the performance and performance of the Dropout method.)