2 ms·
Well, CNNs are already trained on "a large diverse set" of objects, in many different configurations. By the way, something that is not described in the articl
by YeGoblynQueenne 8y ago
Well, CNNs are already trained on "a large diverse set" of objects, in many different configurations.
By the way, something that is not described in the article- the referenced paper also describes experiments where objects where translated inside the source image itself - so not in another image. And the tested networks were still confused.
So it's not "foreign" objects that are the problem. It's that neural nets generalise very poorly. Show them something they haven't seen before and their ability to recognise it falls off a cliff.