Does anyone have any experience/good sources on whether its better to transform multiple class classification problems into a set of binary sub-problems?
For anyone unfamiliar with the terms; a multi-class MLP may predict whether an example is a cat or dog or human; a binary MLP may predict whether an example is a car or not.
When is it better to train a network which predicts if an example is a cat/dog/human vs. training 3 networks which individually predict if an example is cat/not, dog/not, human/not?
[link][1 comment]