I tried the hosted web app for bear classification (https://fastai-v3.onrender.com). I live in Africa and we don’t have bears here. So I tried something else:
That is not a grizzly bear! To improve the bear classifier, we need a “not a bear at all” class. It occurs to me that this must be a general problem with classification - how do you handle the null class? Is is possible/desirable to train a single net to handle this, or must we approach the issue sequentially: net1: bear or not-bear? net2: what kind of bear? (I think the human brain works more like this: Is that an animal? Yes! What kind of animal? Bear! What kind of bear?)