What to do with images that do not fall under the categories in the neural network?

Question

Neural networks

Good afternoon!
Please tell me how best to get around the moment when, for example, an elephant is slipped into a network trained for cats and dogs? In this case, the network will determine it as a cat or a dog, since it has no other option. It is logical that the "other" option is needed, but how to teach it correctly? On what? Or is there another mechanism? Since if there is a grid that defines 100 animals, and the user shoves a photo of a car into it, there will be an incident.
Thank you.

Answer 1

In such cases, the grid trains on three classes: cat, dog, etc. For other things, one of the ready-made datasets is taken or your own is mixed from different sources.
Without a sufficiently large number of training samples, the algorithm is not immune from mistaking an elephant for a cat.
As far as I know, there is no other way to classify.

Answer 2

Nothing...that's the point. If your elephant is 90% (or what is your percentage of coincidence?) Similar to a cat or a dog, then he will attribute them to them. If it doesn't look like it, then put it in a separate place with a label, there is no required percentage of the ratio, then give the choice to the moderator / operator that this is the left image (if required) or what it refers to.
Naturally, with this mark, pictures should not participate in training and setting weights.