Cet article est un compte-rendu informel et approximatif de l’étude publiée dans Bulletin of Sociological Methodology par Antoine Mazières et Camille Roth. Les résultats, l’article, les données et le code sont disponible sur ce site.
Imaginez vous un instant dans un petit village français, en plein moyen-âge. Il y a 5 personnes qui s’appelle Antoine dans le coin et, pour une raison quelconque, vous devez les distinguer les uns des autres, par exemple :
- « Salut ! Je crois que j’ai vu Antoine voler ton scooter hier ? »
- « Quoi !? Quel Antoine ? »
- « Celui avec la maison toute pourrie. …
The episode has become legendary in computer science history. In October 2012 the ECCV conference brought together researchers specialized in computer vision:
So guess who turned up at the 2012 contest? Hinton [the “father” of neural networks revival] and that really shook things up. He didn’t know anything about the field of computer vision, so he took two young guys to change it all! One of them [Alex Krizhevsky] he locked up in a room, telling him: “You can’t come out until it works!” He got huge machines to work, machines that had GPUs which at the time weren’t great, but he got them to communicate with one another to boost them. It was totally crazy computer stuff. Otherwise, it wouldn’t have worked; totally incredible geek knowledge, programming. At the time, computer vision people had been excited about ImageNet for three years [a database of 1.2 million images tagged with 1,000 categories used as a benchmark to compare the classification results of different competitors]. Number 1 had an error rate of 27.03%, number 2 had 27.18%, and number 3 had 27.68%. Hinton sent in this guy from nowhere: “we got a really big deep one to work, we got 17%!” He won over everyone by 10 points! So that young geek did it, and he announced the result in front of the jam-packed room. He didn’t understand anything at all, like he was 17! He didn’t know why those things were there. He’d been locked up in his office and didn’t know anything about the field. And then all of a sudden, there he was in front of Fei-Fei, with LeCun sitting at the back of the room and getting up to answer questions [Li Fei-Fei, professor of computer science and director of SAIL, the Stanford Artificial Intelligence Laboratory; Yann LeCun, today the director of FAIR, Facebook AI Research, and one of the central players in the renewal of neural networks]. And all the big-wigs of computer vision were trying to react: “But that’s not possible. …