2
$\begingroup$

For example, I'm wondering about in this image enter image description here

Why is it that after each convolutional layer, it is preferred that we increase the number of channel (make the blocks longer)? What is the benefit of having multiple output channels rather than just outputting to a single channel?

$\endgroup$
1
  • $\begingroup$ it ends up happending that each channel captures different features from the image like how sharp are the edges, does it have corners, and such. So more channels means capturing more features that might be used to finally be able to classify your output. Not sure what would happen if you merge them all together in a single bigger channel $\endgroup$
    – juvian
    Aug 6 '19 at 15:35
0
$\begingroup$

It's for the network to learn more global features in deeper layers. The example image you have is a classification problem so the network needs to move from local features (from pixels) to global ones (encoded in channels)

This might be a useful visualisation that is also mentioned in the deep learning book: enter image description here

$\endgroup$
2
  • $\begingroup$ But why does having more channels allow for learning of global features? I'm not asking why we have multiple layers; I'm asking why we have multiple output channels that we then feed into fully connected layers at the end, as opposed to having a single output channel that we could then feed into some other layers. $\endgroup$
    – Anon
    Aug 6 '19 at 21:03
  • $\begingroup$ Each channel has it's own filters so different features of the data are learned. Channels is an easy way to structure the network so that multiple representaions are learned from the same input $\endgroup$ Aug 6 '19 at 22:34

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.