The hacky solutions don't really work... the fact is that if you're going to show the model ones and zeros, the model will learn to predict ones and zeros. If the label however would be a probability, then yes it would be a good measure of confidence.
Viewing a single comment thread. View all comments