Dropout Profile
1. In the machine learning model, the parameters of the model too much, too little training samples, trained model is prone to over-fitting phenomenon. Overfitting specific performance: the training data model on a smaller loss function, forecasting
higher accuracy; but in the loss of function test data is relatively large, low prediction accuracy.
2.Dropout: when the forward spread, so that the value of a neuron is activated with a certain probability p to stop working , so you can make more generalized model, because it does not rely too much on some local features.
.Dropout workflow and use
Suppose we want to train such a neural network, shown in Figure 2:
the normal process is: x is the input, the output is y, x we first before propagation through the network, then the error back propagation allows to determine how to update the parameters network learning. Dropout After use, the process becomes as follows:
(1) First, random (provisional) Deletion network half hidden neurons , output neurons remains constant input (broken line in FIG. 3 is a partial temporary neuron deleted)
(2 ) and the input x before propagation through the network modified, then the resulting loss by modifying the back propagation network. After a small number of training samples after executing this process, the neurons are not deleted update a corresponding method parameters (w, b) decreased in accordance with the stochastic gradient.
(3) and then continue to repeat the process (1 and 2):
restore neurons are deleted (for the removed intact neurons, but neurons has been deleted not updated).
Dropout using neural networks
No Dropout network formula:
a network Dropout of formula:
Why Dropout can be solved fit?
（1）取平均的作用： 先回到标准的模型即没有dropout，我们用相同的训练数据去训练5个不同的神经网络，一般会得到5个不同的结果，此时我们可以采用 “5个结果取均值”或者“多数取胜的投票策略”去决定最终结果。例如3个网络判断结果为数字9,那么很有可能真正的结果就是数字9，其它两个网络给出了错误结果。这种“综合起来取平均”的策略通常可以有效防止过拟合问题。因为不同的网络可能产生不同的过拟合，取平均则有可能让一些“相反的”拟合互相抵消。dropout掉不同的隐藏神经元就类似在训练不同的网络，随机删掉一半隐藏神经元导致网络结构已经不同，整个dropout过程就相当于对很多个不同的神经网络取平均。而不同的网络产生不同的过拟合，一些互为“反向”的拟合相互抵消就可以达到整体上减少过拟合。
（2）减少神经元之间复杂的共适应关系： 因为dropout程序导致两个神经元不一定每次都在一个dropout网络中出现。这样权值的更新不再依赖于有固定关系的隐含节点的共同作用，阻止了某些特征仅仅在其它特定特征下才有效果的情况 。迫使网络去学习更加鲁棒的特征 ，这些特征在其它的神经元的随机子集中也存在。换句话说假如我们的神经网络是在做出某种预测，它不应该对一些特定的线索片段太过敏感，即使丢失特定的线索，它也应该可以从众多其它线索中学习一些共同的特征。从这个角度看dropout就有点像L1，L2正则，减少权重使得网络对丢失特定神经元连接的鲁棒性提高。
（3）Dropout类似于性别在生物进化中的角色：物种为了生存往往会倾向于适应这种环境，环境突变则会导致物种难以做出及时反应，性别的出现可以繁衍出适应新环境的变种，有效的阻止过拟合，即避免环境改变时物种可能面临的灭绝。
