本文共 1310 字，大约阅读时间需要 4 分钟。

吴恩达 Andrew Ng

深层神经网络(Deep Neural Networks)

Deep L-layer Neural Networks

• 有一个隐藏层的神经网络，就是一个两层神经网络
• 算神经网络的层数时，只算隐藏层和输出层，输入层看作第0层
• 有一些函数浅层的网络学习不了

Forward propagation in a deep network

Getting your matrix dimensions right

• $W^{[l]}:(n^{[l]},n^{[l-]})\quad b^{[l]}:(n^{[l]},1)$
• $A^{[l]}:(n^{[l]},m)\qquad Z^{[l]}:(n^{[l]},m)$
• 确保前后矩阵一致

Why deep representations?

• 登堂入室
• 边缘探测器针对小区域
• put together simpler things that has detected in order to detect more complex things
• Circuit theory (电路理论)
• There are some functions that you can compute with a L-layer deep neural network while shallower networks require exponentially more hidden units to compute.
• 对于有些函数深层网络更合适

Forward and backward functions

• 

• backward propagation

  

• summary

  

• lots of complexity of your learning algorithm comes from the data rather than your code

Parameters and Hyper parameters

• parameters(参数): W, b
• hyper parameters(超参数) (determines the parameters)
  • learning rate
  • the number of iterations of gradient descent
  • the number of hidden layers
  • choice of activation function
• 不断实验，找到合适的超参数 cycle of (Idea, Code, Experiment)
• just try a few values for the hyper parameters and double check

Deep learning and the human brain

• Today even neuroscientists have almost no idea what even a single neuron is doing.
•