I would answer no, but am not sure if I'm missing something and hope you can help me out:

The derivative of a max-pooling layer in a ConvNet is one w.r.t. the maximum value and zero for all others. A max-pooling layer thus propagates the derivative only to the maximum values. Therefore it sounds like a problem in terms of vanishing gradients similar to the dead ReLU?

On the other hand, a max-pooling layer is acting on the activation maps and we want to update the parameters of the filters generated these activation maps, so the max-pooling layer shouldn't affect the gradient w.r.t. to the filter parameters, right? And thus does not reduce the gradients?

Thank you very much in advance!


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.