This article focus on the back propagate in CNN and its math foundation.
The following picture shows the last 3 layer of CNN. This CNN has n layers totally. And we will demonstrate the back propagate and weight updating process of on it.
Then the Loss function is:
Update the weights of last layer (layer n)
First, let’s update the weights of last layer. That is: update all parameters of the 2 filters of layer n.
So apply chain rule again:
Update the weights of inner layer (eg: layer n-1)
Second, let’s update the weights of inner layers (eg: layer n-1). That is: update all parameters of the 3 filters of layer n-1.
Instead of stopping at equation ③, we have to continue:
The network reaches a convergence when the outcome of Loss function is stably less than a determined value
This is only a study note. Correctness is not guaranteed
Original work. Error correction and Forwarding is welcomed.