神经网络与深度学习笔记(番外)反向传播推导

回顾

我们设 神经网络与深度学习笔记(番外)反向传播推导 - 图5 为第 神经网络与深度学习笔记(番外)反向传播推导 - 图6 层的单元数

则它们的维数

神经网络与深度学习笔记(番外)反向传播推导 - 图7%5C%5C%0A#card=math&code=w%5E%7B%5Bl%5D%7D%2C%20dw%20%EF%BC%9A%28n%5E%7B%5Bl%5D%7D%2Cn%5E%7B%5Bl-1%5D%7D%29%5C%5C%0A)

神经网络与深度学习笔记(番外)反向传播推导 - 图8%5C%5C%0A#card=math&code=b%5E%7B%5Bl%5D%7D%2C%20db%20%EF%BC%9A%28n%5E%7B%5Bl%5D%7D%2C1%29%5C%5C%0A)

神经网络与深度学习笔记(番外)反向传播推导 - 图9%5C%5C%0A#card=math&code=z%5E%7B%5Bl%5D%7D%2Ca%5E%7Bl%7D%3A%28n%5E%7B%5Bl%5D%7D%2C1%29%5C%5C%0A)

神经网络与深度学习笔记(番外)反向传播推导 - 图10%0A#card=math&code=Z%5E%7Bl%7D%2CA%5E%7Bl%7D%2CdZ%2CdA%3A%28n%5E%7B%5Bl%5D%7D%2Cm%29%0A)

反向传播公式为:

神经网络与深度学习笔记(番外)反向传播推导 - 图11%5C%5C%0A#card=math&code=dz%5E%7B%5Bl%5D%7D%20%3D%20da%5E%7B%5Bl%5D%7D%20%2A%20g%5E%7B%5Bl%5D%27%7D%28z%5E%7B%5Bl%5D%7D%29%5C%5C%0A)

神经网络与深度学习笔记(番外)反向传播推导 - 图12

神经网络与深度学习笔记(番外)反向传播推导 - 图13

神经网络与深度学习笔记(番外)反向传播推导 - 图14

推导

首先我们知道

神经网络与深度学习笔记(番外)反向传播推导 - 图15

神经网络与深度学习笔记(番外)反向传播推导 - 图16%0A#card=math&code=a%5E%7B%5Bl%5D%7D%20%3D%20g%5E%7B%5Bl%5D%7D%28z%5E%7B%5Bl%5D%7D%29%0A)

神经网络与深度学习笔记(番外)反向传播推导 - 图17%20%3D%20-%20yloga-(1-y)log(1-a)%5C%5C%0A#card=math&code=%5Cjmath%28a%2Cy%29%20%3D%20-%20yloga-%281-y%29log%281-a%29%5C%5C%0A)

接下来开始推导过程:

神经网络与深度学习笔记(番外)反向传播推导 - 图18的证明

由公式 神经网络与深度学习笔记(番外)反向传播推导 - 图19%20%3D%20-%20yloga-(1-y)log(1-a)%5C%5C#card=math&code=%5Cjmath%28a%2Cy%29%20%3D%20-%20yloga-%281-y%29log%281-a%29%5C%5C) 对 神经网络与深度学习笔记(番外)反向传播推导 - 图20 求导得:

神经网络与深度学习笔记(番外)反向传播推导 - 图21%7D%7Bda%5E%7B%5Bl%5D%7D%7D%20%3D%20-%5Cfrac%7By%7D%7Ba%5E%7B%5Bl%5D%7D%7D%20%2B%20%5Cfrac%7B1-y%7D%7B1-a%5E%7B%5Bl%5D%7D%7D%5C%5C%0A#card=math&code=%5Cfrac%7Bd%5Cjmath%28a%5E%7B%5Bl%5D%7D%2Cy%29%7D%7Bda%5E%7B%5Bl%5D%7D%7D%20%3D%20-%5Cfrac%7By%7D%7Ba%5E%7B%5Bl%5D%7D%7D%20%2B%20%5Cfrac%7B1-y%7D%7B1-a%5E%7B%5Bl%5D%7D%7D%5C%5C%0A)

神经网络与深度学习笔记(番外)反向传播推导 - 图22%7D%7Bdz%5E%7B%5Bl%5D%7D%7D%20%3D%20%5Cfrac%7Bd%5Cjmath(a%5E%7B%5Bl%5D%7D%2Cy)%7D%7Bda%5E%7B%5Bl%5D%7D%7D*%5Cfrac%7Bda%5E%7B%5Bl%5D%7D%7D%7Bdz%5E%7B%5Bl%5D%7D%7D%0A#card=math&code=%5Cfrac%7Bd%5Cjmath%28a%5E%7B%5Bl%5D%7D%2Cy%29%7D%7Bdz%5E%7B%5Bl%5D%7D%7D%20%3D%20%5Cfrac%7Bd%5Cjmath%28a%5E%7B%5Bl%5D%7D%2Cy%29%7D%7Bda%5E%7B%5Bl%5D%7D%7D%2A%5Cfrac%7Bda%5E%7B%5Bl%5D%7D%7D%7Bdz%5E%7B%5Bl%5D%7D%7D%0A)

神经网络与深度学习笔记(番外)反向传播推导 - 图23%0A#card=math&code=%5Cfrac%7Bda%5E%7B%5Bl%5D%7D%7D%7Bdz%5E%7B%5Bl%5D%7D%7D%20%3D%20g%5E%7B%5Bl%5D%27%7D%28z%5E%7B%5Bl%5D%7D%29%0A)

所以代入公式 中得:

神经网络与深度学习笔记(番外)反向传播推导 - 图24%7D%7Bdz%5E%7B%5Bl%5D%7D%7D%20%3Dda%5E%7B%5Bl%5D%7D*g%5E%7B%5Bl%5D’%7D(z%5E%7B%5Bl%5D%7D)%0A#card=math&code=%5Cfrac%7Bd%5Cjmath%28a%5E%7B%5Bl%5D%7D%2Cy%29%7D%7Bdz%5E%7B%5Bl%5D%7D%7D%20%3Dda%5E%7B%5Bl%5D%7D%2Ag%5E%7B%5Bl%5D%27%7D%28z%5E%7B%5Bl%5D%7D%29%0A)

注意:上式子中的 神经网络与深度学习笔记(番外)反向传播推导 - 图25 为简写,实际上是:

神经网络与深度学习笔记(番外)反向传播推导 - 图26%7D%7Bda%5E%7B%5Bl%5D%7D%7D%0A#card=math&code=%5Cfrac%7Bd%5Cjmath%28a%5E%7B%5Bl%5D%7D%2Cy%29%7D%7Bda%5E%7B%5Bl%5D%7D%7D%0A)

后面的神经网络与深度学习笔记(番外)反向传播推导 - 图27 同理

即证明了公式 神经网络与深度学习笔记(番外)反向传播推导 - 图28%5C%5C#card=math&code=dz%5E%7B%5Bl%5D%7D%20%3D%20da%5E%7B%5Bl%5D%7D%20%2A%20g%5E%7B%5Bl%5D%27%7D%28z%5E%7B%5Bl%5D%7D%29%5C%5C)

神经网络与深度学习笔记(番外)反向传播推导 - 图29的证明

神经网络与深度学习笔记(番外)反向传播推导 - 图30%7D%7Bdz%5E%7B%5Bl%5D%7D%7D%20%3Dda%5E%7B%5Bl%5D%7D*g%5E%7B%5Bl%5D’%7D(z%5E%7B%5Bl%5D%7D)%0A#card=math&code=%5Cfrac%7Bd%5Cjmath%28a%5E%7B%5Bl%5D%7D%2Cy%29%7D%7Bdz%5E%7B%5Bl%5D%7D%7D%20%3Dda%5E%7B%5Bl%5D%7D%2Ag%5E%7B%5Bl%5D%27%7D%28z%5E%7B%5Bl%5D%7D%29%0A)

的证明结果,我们来推一下 神经网络与深度学习笔记(番外)反向传播推导 - 图31

神经网络与深度学习笔记(番外)反向传播推导 - 图32%7D%7Bdw%5E%7B%5Bl%5D%7D%7D%20%3D%5Cfrac%7Bd%5Cjmath(a%5E%7B%5Bl%5D%7D%2Cy)%7D%7Bdz%5E%7B%5Bl%5D%7D%7D%20*%20%5Cfrac%7Bdz%5E%7B%5Bl%5D%7D%7D%7Bdw%5E%7B%5Bl%5D%7D%7D%0A#card=math&code=%5Cfrac%7Bd%5Cjmath%28a%5E%7B%5Bl%5D%7D%2Cy%29%7D%7Bdw%5E%7B%5Bl%5D%7D%7D%20%3D%5Cfrac%7Bd%5Cjmath%28a%5E%7B%5Bl%5D%7D%2Cy%29%7D%7Bdz%5E%7B%5Bl%5D%7D%7D%20%2A%20%5Cfrac%7Bdz%5E%7B%5Bl%5D%7D%7D%7Bdw%5E%7B%5Bl%5D%7D%7D%0A)

因为:

神经网络与深度学习笔记(番外)反向传播推导 - 图33

所以:

神经网络与深度学习笔记(番外)反向传播推导 - 图34

故:

神经网络与深度学习笔记(番外)反向传播推导 - 图35%7D%7Bdw%5E%7B%5Bl%5D%7D%7D%20%3D%5Cfrac%7Bd%5Cjmath(a%5E%7B%5Bl%5D%7D%2Cy)%7D%7Bdz%5E%7B%5Bl%5D%7D%7D%20%20%5Cfrac%7Bdz%5E%7B%5Bl%5D%7D%7D%7Bdw%5E%7B%5Bl%5D%7D%7D%3Ddz%5E%7B%5Bl%5D%7Da%5E%7B%5Bl-1%5D%7D%0A#card=math&code=%5Cfrac%7Bd%5Cjmath%28a%5E%7B%5Bl%5D%7D%2Cy%29%7D%7Bdw%5E%7B%5Bl%5D%7D%7D%20%3D%5Cfrac%7Bd%5Cjmath%28a%5E%7B%5Bl%5D%7D%2Cy%29%7D%7Bdz%5E%7B%5Bl%5D%7D%7D%20%2A%20%5Cfrac%7Bdz%5E%7B%5Bl%5D%7D%7D%7Bdw%5E%7B%5Bl%5D%7D%7D%3Ddz%5E%7B%5Bl%5D%7D%2Aa%5E%7B%5Bl-1%5D%7D%0A)

神经网络与深度学习笔记(番外)反向传播推导 - 图36 的证明

神经网络与深度学习笔记(番外)反向传播推导 - 图37%7D%7Bdz%5E%7B%5Bl%5D%7D%7D%20%3Dda%5E%7B%5Bl%5D%7D*g%5E%7B%5Bl%5D’%7D(z%5E%7B%5Bl%5D%7D)%0A#card=math&code=%5Cfrac%7Bd%5Cjmath%28a%5E%7B%5Bl%5D%7D%2Cy%29%7D%7Bdz%5E%7B%5Bl%5D%7D%7D%20%3Dda%5E%7B%5Bl%5D%7D%2Ag%5E%7B%5Bl%5D%27%7D%28z%5E%7B%5Bl%5D%7D%29%0A)

的证明结果,我们来推一下 神经网络与深度学习笔记(番外)反向传播推导 - 图38

神经网络与深度学习笔记(番外)反向传播推导 - 图39%7D%7Bdb%5E%7B%5Bl%5D%7D%7D%20%3D%5Cfrac%7Bd%5Cjmath(a%5E%7B%5Bl%5D%7D%2Cy)%7D%7Bdz%5E%7B%5Bl%5D%7D%7D%20*%20%5Cfrac%7Bdz%5E%7B%5Bl%5D%7D%7D%7Bdb%5E%7B%5Bl%5D%7D%7D%0A#card=math&code=%5Cfrac%7Bd%5Cjmath%28a%5E%7B%5Bl%5D%7D%2Cy%29%7D%7Bdb%5E%7B%5Bl%5D%7D%7D%20%3D%5Cfrac%7Bd%5Cjmath%28a%5E%7B%5Bl%5D%7D%2Cy%29%7D%7Bdz%5E%7B%5Bl%5D%7D%7D%20%2A%20%5Cfrac%7Bdz%5E%7B%5Bl%5D%7D%7D%7Bdb%5E%7B%5Bl%5D%7D%7D%0A)

因为:

神经网络与深度学习笔记(番外)反向传播推导 - 图40

所以:

神经网络与深度学习笔记(番外)反向传播推导 - 图41

故:

神经网络与深度学习笔记(番外)反向传播推导 - 图42%7D%7Bdb%5E%7B%5Bl%5D%7D%7D%20%3D%5Cfrac%7Bd%5Cjmath(a%5E%7B%5Bl%5D%7D%2Cy)%7D%7Bdz%5E%7B%5Bl%5D%7D%7D%20*%20%5Cfrac%7Bdz%5E%7B%5Bl%5D%7D%7D%7Bdb%5E%7B%5Bl%5D%7D%7D%3Ddz%5E%7B%5Bl%5D%7D%0A#card=math&code=%5Cfrac%7Bd%5Cjmath%28a%5E%7B%5Bl%5D%7D%2Cy%29%7D%7Bdb%5E%7B%5Bl%5D%7D%7D%20%3D%5Cfrac%7Bd%5Cjmath%28a%5E%7B%5Bl%5D%7D%2Cy%29%7D%7Bdz%5E%7B%5Bl%5D%7D%7D%20%2A%20%5Cfrac%7Bdz%5E%7B%5Bl%5D%7D%7D%7Bdb%5E%7B%5Bl%5D%7D%7D%3Ddz%5E%7B%5Bl%5D%7D%0A)

神经网络与深度学习笔记(番外)反向传播推导 - 图43 的证明

神经网络与深度学习笔记(番外)反向传播推导 - 图44%7D%7Bdz%5E%7B%5Bl%5D%7D%7D%20%3Dda%5E%7B%5Bl%5D%7D*g%5E%7B%5Bl%5D’%7D(z%5E%7B%5Bl%5D%7D)%0A#card=math&code=%5Cfrac%7Bd%5Cjmath%28a%5E%7B%5Bl%5D%7D%2Cy%29%7D%7Bdz%5E%7B%5Bl%5D%7D%7D%20%3Dda%5E%7B%5Bl%5D%7D%2Ag%5E%7B%5Bl%5D%27%7D%28z%5E%7B%5Bl%5D%7D%29%0A)

的证明结果,我们来推一下 神经网络与深度学习笔记(番外)反向传播推导 - 图45

神经网络与深度学习笔记(番外)反向传播推导 - 图46%7D%7Bda%5E%7B%5Bl-1%5D%7D%7D%20%3D%5Cfrac%7Bd%5Cjmath(a%5E%7B%5Bl%5D%7D%2Cy)%7D%7Bdz%5E%7B%5Bl%5D%7D%7D%20*%20%5Cfrac%7Bdz%5E%7B%5Bl%5D%7D%7D%7Bda%5E%7B%5Bl-1%5D%7D%7D%0A#card=math&code=%5Cfrac%7Bd%5Cjmath%28a%5E%7B%5Bl%5D%7D%2Cy%29%7D%7Bda%5E%7B%5Bl-1%5D%7D%7D%20%3D%5Cfrac%7Bd%5Cjmath%28a%5E%7B%5Bl%5D%7D%2Cy%29%7D%7Bdz%5E%7B%5Bl%5D%7D%7D%20%2A%20%5Cfrac%7Bdz%5E%7B%5Bl%5D%7D%7D%7Bda%5E%7B%5Bl-1%5D%7D%7D%0A)

因为:

神经网络与深度学习笔记(番外)反向传播推导 - 图47

所以:

神经网络与深度学习笔记(番外)反向传播推导 - 图48

故:

神经网络与深度学习笔记(番外)反向传播推导 - 图49%7D%7Bda%5E%7B%5Bl-1%5D%7D%7D%20%3D%5Cfrac%7Bd%5Cjmath(a%5E%7B%5Bl%5D%7D%2Cy)%7D%7Bdz%5E%7B%5Bl%5D%7D%7D%20%20%5Cfrac%7Bdz%5E%7B%5Bl%5D%7D%7D%7Bda%5E%7B%5Bl-1%5D%7D%7D%3Dw%5E%7B%5Bl%5D%5E%7BT%7D%7Ddz%5E%7B%5Bl%5D%7D%0A#card=math&code=%5Cfrac%7Bd%5Cjmath%28a%5E%7B%5Bl%5D%7D%2Cy%29%7D%7Bda%5E%7B%5Bl-1%5D%7D%7D%20%3D%5Cfrac%7Bd%5Cjmath%28a%5E%7B%5Bl%5D%7D%2Cy%29%7D%7Bdz%5E%7B%5Bl%5D%7D%7D%20%2A%20%5Cfrac%7Bdz%5E%7B%5Bl%5D%7D%7D%7Bda%5E%7B%5Bl-1%5D%7D%7D%3Dw%5E%7B%5Bl%5D%5E%7BT%7D%7D%2Adz%5E%7B%5Bl%5D%7D%0A)

会不会觉得很奇怪?

为什么 神经网络与深度学习笔记(番外)反向传播推导 - 图50 是个转置?

神经网络与深度学习笔记(番外)反向传播推导 - 图51%5C%5C%0A#card=math&code=w%5E%7B%5Bl%5D%7D%2C%20dw%20%EF%BC%9A%28n%5E%7B%5Bl%5D%7D%2Cn%5E%7B%5Bl-1%5D%7D%29%5C%5C%0A)

神经网络与深度学习笔记(番外)反向传播推导 - 图52%5C%5C%0A#card=math&code=b%5E%7B%5Bl%5D%7D%2C%20db%20%EF%BC%9A%28n%5E%7B%5Bl%5D%7D%2C1%29%5C%5C%0A)

神经网络与深度学习笔记(番外)反向传播推导 - 图53%5C%5C%0A#card=math&code=z%5E%7B%5Bl%5D%7D%2Ca%5E%7Bl%7D%3A%28n%5E%7B%5Bl%5D%7D%2C1%29%5C%5C%0A)

因为我们还要考虑维度的问题,

神经网络与深度学习笔记(番外)反向传播推导 - 图54 维度为 神经网络与深度学习笔记(番外)反向传播推导 - 图55#card=math&code=%28n%5E%7B%5Bl%5D%7D%2Cn%5E%7B%5Bl-1%5D%7D%29)

神经网络与深度学习笔记(番外)反向传播推导 - 图56维度为神经网络与深度学习笔记(番外)反向传播推导 - 图57#card=math&code=%28n%5E%7B%5Bl%5D%7D%2C1%29)

神经网络与深度学习笔记(番外)反向传播推导 - 图58 维度为 神经网络与深度学习笔记(番外)反向传播推导 - 图59#card=math&code=%28n%5E%7B%5Bl%5D%7D%2C1%29)

故要使得 神经网络与深度学习笔记(番外)反向传播推导 - 图60神经网络与深度学习笔记(番外)反向传播推导 - 图61 的积等于 神经网络与深度学习笔记(番外)反向传播推导 - 图62 ,我们需要将神经网络与深度学习笔记(番外)反向传播推导 - 图63 转置,转置后的 神经网络与深度学习笔记(番外)反向传播推导 - 图64 维度为 神经网络与深度学习笔记(番外)反向传播推导 - 图65#card=math&code=%28n%5E%7B%5Bl%5D%7D%2Cn%5E%7B%5Bl-1%5D%7D%29) ,才可以使得等式成立,且维度一致。