Here's the famous VGG-16 model.


Do the inputs and outputs of a convolutional layer, before pooling, usually have the same depth? What's the reason for that?

Is there a theory or paper trying to explain this kind of setting?

  • $\begingroup$ I edited this post in order to save it. It wasn't clear what you meant by "input/output channels". To answer what I think is your question: no, the depth of the inputs and outputs of a convolutional layer are not typically the same. $\endgroup$ – nbro Jul 4 at 20:32

Keeping the same channel size allows the model to maintain rank but i would say the main reason is convenience. Its easier book keeping.

Also in many model cases output features need some form of alignment with the input (example being all models using residual units -- $\hat{x} = F(x) + x$

| improve this answer | |
  • $\begingroup$ What do you mean by "rank" in this context? It is also not clear the question because the depth of the input and output volumes are usually different. $\endgroup$ – nbro Jun 10 '19 at 17:11
  • $\begingroup$ also good point about alignment because GPU and CPU vector ops need this $\endgroup$ – user8426627 Jun 10 '19 at 17:54

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.