I don’t think you have to train a model to recognize “not-bearness”. You have to train a model to find features that indicate “bearness” and in the absence of “bearness” it would output “not bear”.
For example I can train a model to recognize my face. I don’t have to give it the face of all other people on earth in order to train it.
As we will see later on, it won’t work with this model which has to produce numbers that add up to one. You need another kinds of model/loss function for this. Stay tune for the lesson about multi-labels problems
So, if face training with pictures of only one face, how to confirm it won’t get confused when tested with different faces? Wouldn’t it need to be trained by finding the difference between faces?
It couldn’t be trained solely on one person’s face. It would have to be a dataset that contains many pictures of the person you’re trying to identify and many pictures of other random people. My point was just that you don’t have to show it every other person in order to build a classifier that works.
In the example of data shift where the classifier goes from bears to raccoons should you do transfer learning from your previous model or retrain completely?