I respectifully disagree. When it comes to uncertainty modelling, what we usually mean (and want) is to have our model be able to output "I have no idea".

For instance, a dogs 'n cats classifier will tell you how sure it is an image is a dog or a cat. However, if you show it frog/banana/elephant images, it will also tell you "its a dog/cat". In other words, it has no "thid option" saying "This input means nothing to me"

In an ideal world, it would output a 50% dog 50% cat when it is unsure, but there is no guarantee it will ever (or reliably) do so, as it might just find frog eyes very cat-ish. Thus, you don't really have uncertainty in a usable sense.

In practical settings, uncertainty modelling is used when you need models to predict stuff but the domain is large and wrong predictions are costly. For instance, an alarm system. There are MANY things that can show up on surveillance cameras that could trick a CNN into firing the alarm. Using uncertainty modelling, you hope to curb false-positives by making sure the alarm will only fire to situations that it was trained upon, not to anything take trickles its matrices.

Therefore, just comparing class scores isn't really a good idea. That's why you have techniques like monte carlo dropout to sample how sensitive the neural network is to the input.