Question 5/13 fast.ai v3 lecture 11
What is one way of handling a scenario where your model, during inference, might be run on instances of classes not included in the train set?
If you foresee this situation occuring, you could create a class 'OTHER' and populate it with something that would make sense, that could mimic well what you expect your model to exprience at test time. If your training set contains a lot of classes, good candidates for combining into the 'OTHER' class might be members of the least represented categories.
Relevant part of lecture