# Stanford Machine Learning 公开课笔记(3) Neural Network

1 如何决定神经网络有多少层？
3 神经网络和逻辑回归什么关系?

Why do we Need Neural Networks?
because we sometimes need to learn complex machine learning hypothesis.如下图

Neural network model representation-1
The leftmost is the input layer, the rightmost layer is the output layer. the rest are the hidden layers.

Neural network model representation-2

Neural network model representation-3

How can neural network handle complex non-linear questions? Intuition1

Neural Network Cost Function

2. 简化问题，假设输入只有一份，用这一份输入进行训练。

What these delta terms are？they turn out to be the partial derivative of the cost function with respect to these intermediate terms that we're computing.
And so their measure of how much would we like to change the neural network's weights in order to
affect these intermediate values of the computation,
so as to affect the final output the neural network h of x and therefore affect the overall cost.

Octave中使用Neural Network的注意事项1

step1

Octave中使用Neural Network的注意事项2

Octave中使用Neural Network的注意事项3
Random Initialization

Octave中使用Neural Network的注意事项4

The more hidden unit, the better, but to many hidden units might be computationally expensive.

Training的过程步骤

1 如何决定神经网络有多少层？
理论上说，层数越多，效果越好，但是开销也会越高。实际用途中，一般选择使用hidden layer1~2层。
直接使用对矩阵求导的数学公式。
3 神经网络和逻辑回归什么关系?
可以认为神经网络把逻辑回归进行了很多层，每一层的输出都是下一层的输入。

• 广告
• 抄袭
• 版权
• 政治
• 色情
• 无意义
• 其他

120