【读书1】【2017】MATLAB与深度学习——单层神经网络的训练：增量规则(2)

5. 对所有训练数据执行第2 – 4步。

5. Perform Steps 2-4 for all training data.

6. 重复第2 – 5步，直到误差达到可接受的水平。

7. Repeat Steps 2-5 until the error reachesan acceptable tolerance level.

这些步骤与“神经网络的监督学习”部分中的监督学习过程几乎相同。

These steps are almost identical to that ofthe process for the supervised learning in the “Supervised Learning of a NeuralNetwork” section.

唯一的区别是增加了步骤6。

The only difference is the addition of Step6.

第6步只是说明整个训练过程的重复。

Step 6 just states that the whole trainingprocess is repeated.

一旦步骤5已经完成，模型已经被训练到每个数据点。

Once Step 5 has been completed, the modelhas been trained with every data point.

那么，为什么我们要使用所有相同的训练数据来训练它呢？

Then, why do we train it using all of thesame training data?

这是因为增量规则在重复该过程时搜索解决方案，而不是立即解决所有问题。

This is because the delta rule searches forthe solution as it repeats the process, rather than solving it all at once.

增量规则是一种称为梯度下降的数值方法。

The delta rule is a type of numericalmethod called gradient descent.

梯度下降从初始值开始，并直到求解结束。

The gradient descent starts from theinitial value and proceeds to the solution.

它的名字起源于它的行为，即它在寻找解决方案时，就像一个球沿着最陡的路径滚下山一样。

Its name originates from its behaviorwhereby it searches for the solution as if a ball rolls down the hill along thesteepest path.

在这个类比中，球的位置是模型的偶然输出，而最终的底部是解决方案。

In this analogy, the position of the ballis the occasional output from the model, and the bottom is the solution.

值得注意的是，梯度下降法不能仅用一次投掷就将球扔到底部。

It is noteworthy that the gradient descentmethod cannot drop the ball to the bottom with just one throw.

重复整个过程，对相同的数据再次训练模型可以不断改进模型。

The whole process repeats, as retrainingthe model with the same data may improve the model.

仅供参考，每训练迭代一次，其中所有训练数据都将步骤2 – 5执行一次，该过程也称为纪元或时代。

Just for reference, the number of trainingiterations, in each of which all training data goes through Steps 2-5 once, iscalled an epoch.

——本文译自Phil Kim所著的《Matlab Deep Learning》

更多精彩文章请关注微信号：