Once the model is trained, we use the same features to get our answer. Of course, with the question we asked earlier, if we feed in images of cats, we will get a positive answer. But if we feed in an image of a tiger, a lion, or a dog, we will also get a positive identification. So the question we asked is not, "Are these images of cats?", but really, "Are these images of cats, knowing that cats have two ears, two eyes, a nose, a mouth, and whiskers?". Our definition of a cat was wrong and lead us to wrong answers.

This is where know-how and practice are important. Designing the right model to answer the question you have been asked is something that anyone can do once this important point has been understood.