I know back propagation takes derivatives (changing one quantity wrt other). But how this is applied when there is maxpooling layer in between two Conv2D layers? How it gains its original shape when there is padding added?
0
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
Max pooling will cancel the effect of not pooled values to the gradients. Padded values either have no effect.
Nice thing about convolution is, that it is basically reducable to a matrix multiplication and the backpropagation is simply the transposed of it. So you have already your backward pass stored in the foward pass.
