0
$\begingroup$

In the paper by Kendall and Gal (What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?), dropout is being set after every convolutional layer. However, is it still legit to estimate epistemic and aleatoric uncertainty on the same way but using different architecture that doesn't have dropout after every weight layer but rather putting it only on FC layers in the network?

$\endgroup$

1 Answer 1

1
$\begingroup$

Well, "legit" is not a thing in deep learning. This is an empirical field, and we use whatever works. There are no deep learning police who are going to tell you that you are doing it wrong.

I suspect you are asking whether the method will still be as effective if you only put dropout on the fully connected layers. The way to figure that out is to try both and compare how well each works. I wouldn't be surprised if it works less well to have dropout after only the fully connected layers, but it's hard to know without trying it.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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