我们可以将它们与D给出的正确输出进行比较。
We can compare them with the correctoutputs given by D.
究竟发生了什么呢?
What happened?
我们得到两组完全不同的数据集合。
We got two totally different sets.
即使长时间训练神经网络,输出结果也并没有什么变化。
Training the neural network for a longerperiod does not make a difference.
与“示例:增量规则”一节中代码唯一的区别是正确输出变量D发生了变化。
The only difference from the code from the“Example: Delta Rule” section is the correct output variable, D.
究竟发生了什么呢?
What actually happened?
通过分析训练数据有助于说明这个问题。
Illustrating the training data can helpelucidate this problem.
让我们把输入数据的三个值分别解释为X、Y和Z坐标。
Let’s interpret the three values of theinput data as the X, Y, and Z coordinates, respectively.
由于第三个值,即Z坐标被固定为1,训练数据可以在平面上可视化,如图2-22所示。
As the third value, i.e. the Z coordinate,is fixed as 1, the training data can be visualized on a plane as shown inFigure 2-22.
图2-22 将输入数据的三个值解释为X、Y和Z坐标Interpretingthe three values of the input data as the X, Y, and Z coordinates
圆圈中的值0和1是分配给每个点的正确输出。
The values 0 and 1 in the circles are the correct outputs assigned to each point.
从上图中可以看出,我们不能用直线来划分0和1的区域。
One thing to notice from this figure isthat we cannot divide the regions of 0 and 1 with a straight line.
但是,我们可以用一条复杂的曲线将0和1的区域分开,如图2-23所示。
However, we may divide it with acomplicated curve, as shown in Figure 2-23.
图2-23 我们只能用复杂的曲线把0和1的区域分开We can only separate the regions of 0 and 1 with a complicated curve
这种类型的问题被称为是线性不可分割的。
This type of problem is said to be linearlyinseparable.
重复以上相同的过程,在X-Y平面上将“示例:增量规则”一节中的训练数据绘制在图2-24中。
In the same process, the training data fromthe “Example: Delta Rule” section on the X-Y plane appears in Figure 2-24.
——本文译自Phil Kim所著的《Matlab Deep Learning》
更多精彩文章请关注微信号: