3
$\begingroup$

How do we know how many feature maps are needed in the convolution layer?

All steps are clear to me except of this step.

$\endgroup$
1
1
$\begingroup$

I'm no expert, but I believe the idea is that each feature map is being trained to capture a particular aspect of the input. So the number of feature maps is really a function of how many different features need to be captured and "understood" by the CNN to produce the desired output.

So, I would suggest that if you are classifying many different things (say, 100 different types of pictures) or if what you are classifying has many different subtle aspects to it that could be used to make the classification differentiation then you would want many feature maps.

However, that will slow things down, so if you're only classifying simple things or have only a few classification types then you could get away with fewer feature maps.

Robert

$\endgroup$

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.