1
$\begingroup$

Why is backpropagation through maxpool and relu needed?

Purpose of backpropagation is to update weights while on the other hand maxpool and relu only perform a simple operation on the input. They don't really have any weights or any filters. So why is there a need to backpropagate through them?

$\endgroup$
  • $\begingroup$ But how can we disconnect the flow in between? Like backprop is Mathematically transfer of gradients all the way back to the input.. $\endgroup$ – Aditya Aug 11 '18 at 23:43
1
$\begingroup$

Why is backpropagation through maxpool and relu needed?

Any differentiable function (max pool and relu) through which inputs pass will have a gradient.. The nonlinear functions are still functions, so the chain rule still applies here...and hence we have the Gradients flowing through them..This gradient will have to be backpropagated...

|improve this answer|||||
$\endgroup$

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.