【读书1】【2017】MATLAB与深度学习——神经网络与分类(1)

比如根据年龄和教育水平进行收入预测分析。

It can be exemplified with the predictionof income for a given age and education level.

虽然神经网络适用于分类和回归，但却很少用于回归。

Although the neural network is applicableto both classification and regression, it is seldom used for regression.

这不是因为它产生了较差的性能，而是因为大多数回归问题可以用更简单的模型来解决。（回归问题过于简单，采用神经网络通常是大材小用了！！！）

This is not because it yields poorperformance, but because most of regression problems can be solved usingsimpler models.

因此，本书主要讲解分类的话题。

Therefore, we will stick to classificationthroughout this book.

在将神经网络应用于分类时，输出层的公式通常根据数据应该分成多少类别而不同。

In the application of the neural network toclassification, the output layer is usually formulated differently depending onhow many groups the data should be divided into.

当使用更多的类别时，节点数目和适当激活函数的选择与只有两类数据的分类是不同的。

The selection of the number of nodes andsuitable activation functions for the classification of two groups is differentwhen using more groups.

请记住，数据分类的类别数量只影响输出节点，而与隐藏节点无关。

Keep in mind that it affects only theoutput nodes, while the hidden nodes remain intact.

当然，本章介绍的方法并不是所有问题都适用的。

Of course, the approaches of this chapterare not only ones available.

然而，这些方法可能是初学者最好的学习材料，因为它们已经经过许多研究和案例验证。

However, these may be the best to startwith, as they have been validated through many studies and cases.

二元分类（Binary Classification）

我们将从二元分类的神经网络开始学习，它的目标是将输入数据分为两类。

We will start with the binaryclassification neural network, which classifies the input data into one of thetwo groups.

这种分类器实际上比你所期望的更有用。

This kind of classifier is actually usefulfor more applications than you may expect.

一些典型的应用包括垃圾邮件过滤（垃圾邮件或普通邮件）和贷款审批（批准或拒绝）。

Some typical applications include spam mailfiltering (a spam mail or a normal mail) and loan approvals (approve or deny).

对于二元分类，单个输出节点的神经网络已经足够了。

For binary classification, a single outputnode is sufficient for the neural network.

这是因为输入数据可以通过神经网络的输出值大于或小于阈值来分类。

This is because the input data can beclassified by the output value, which is either greater than or less than thethreshold.

例如，如果采用sigmoid函数作为输出节点的激活函数，那么可以根据输出是否大于0.5对数据进行分类。

For example, if the sigmoid function isemployed as the activation function of the output node, the data can beclassified by whether the output is greater than 0.5 or not.

由于sigmoid函数的取值范围是0 - 1，所以我们可以在中间位置对数据进行分类，如图4-1所示。

As the sigmoid function ranges from 0-1, wecan divide groups in the middle, as shown in Figure 4-1.

图4-1 二元分类问题Binary classification problem

考虑图4-1所示的二元分类问题。

Consider the binary classification problemshown in Figure 4-1.

对于给定的坐标（x，y），通过该模型确定输入数据属于哪个类别。

For the given coordinates (x, y), the modelis to determine which group the data belongs.

在这种情况下，训练数据以图4-2所示的格式给出。

In this case, the training data is given inthe format shown in Figure 4-2.

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

更多精彩文章请关注微信号：