Question 22/30 v3 lecture 10

For classification, what is a good way of handling items in your dataset that don't belong to any of the classes?


Treat the presence of a class as a binomial - all values below threshold equal no class

Relevant part of lecture

supplementary material

As explained in the lecture, to be able to successfully predict MISSING some activations in the penultimate layer would need to correspond to (grow large for) features detecting 'not-cat', 'not-dog', etc. But in reality there is no set of features that if they are detected it is clearly not a dog or not a cat in an image.