I don't think you read the article very far. Yes, he started with an averaged face (which then matched nothing, by the way). Then he added (I'll simplify for yo

I don't think you read the article very far. Yes, he started with an averaged face (which then matched nothing, by the way). Then he added (I'll simplify for you) random dots to it until it matches something. Thus it is a reverse lookup from the way facial recognition is usually used. Usually you give it a face and see if it has a match. In this case you basically feed it random face-like data until you trigger it, then you have some insight into what it is looking for. The fact that the results look very little like human faces do provide insight most people don't have into how the algorithms work. The algorithms basically work by measuring distances between dots in the end anyway, distance between eyes, etc., and that's very unlike how humans recognize faces which has more depth to it.