1
$\begingroup$

Kaparthy in his blog post said

[this] hints at the kinds of architectures we’ll eventually explore. As an example - are very local features enough or do we need global context?

I'd like to gain expertise in designing networks for local vs global features. Obviously increasing the receptive field of neurons (though more pooling/striding and bigger kernels) will allow to network to take more global context into account.

Could someone point me in the direction of some readings on the topic?

Are there any specific architectures targeted to local features or to global context?

$\endgroup$
1
$\begingroup$

Well, in regards to properties of CNNs in regards to local versus global features, you should familiarize yourself with the concepts of invariance and equivariance. At some point you should also learn about the Picasso problem which is a consequence of the invariances and equivariances of CNNs + pooling.

That will probably also mean you'll encounter Capsule Nets at some point. While not really CNNs, learning about CapsNets can elucidate the weaknesses of convolutions since CapsNets were made to solve several of those.

There are CNN architectures that, in parallel, use different scales of local features, such as the Inception architecture and ResNext; Both combine local features on different scales, i.e. they use differently sized kernels in parallel to improve classifications.

$\endgroup$

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.