We know we've succeeded when the machine can recognize a concept on a regular basis.
And of course, machines might not get it right at first. But we'll keep showing them images of cats and testing them until they do.