3 ms·
I am not sure this works or that success rates like claimed are possible: > In 70 percent of cases, he was quite easily able to identify the people he had phot
by gwern 11y ago
I am not sure this works or that success rates like claimed are possible:
> In 70 percent of cases, he was quite easily able to identify the people he had photographed.
VKontakte has 280m+ accounts and is the second most popular website in Russia (https://en.wikipedia.org/wiki/VK_%28social_networking%29); https://en.wikipedia.org/wiki/VK_%28social_networking%29); the Russian population is ~144m. I can't find any decent estimate of how many Russians use VK, but it seems to be at least as dominate in Russia as Facebook in the USA, which is to say, ~71% (http://www.pewinternet.org/fact-sheets/social-networking-fact-sheet/ http://www.pewinternet.org/fact-sheets/social-networking-fac...). So from non-VK usage alone, he should be able to identify no more than 70-80 percent since you can't identify those who aren't on VK in the first place.
Facial recognition is a notoriously difficult machine learning task, and doing facial recognition on the VK population of St Petersburg (>249688 * 0.7 = >174781) will be challenging even with extremely good facial recognition. I don't know what this 'FindFace' application is, but I'll be charitable and assume it matches the 2015 state of the art, "Facenet: A unified embedding for face recognition and clustering" http://arxiv.org/abs/1503.03832 http://arxiv.org/abs/1503.03832 AFAIK, with a best-case binary classification accuracy of 99.63% (https://en.wikipedia.org/wiki/Accuracy_and_precision#In_binary_classification https://en.wikipedia.org/wiki/Accuracy_and_precision#In_bina...):
0.9963 = true positives + true negatives / true positives + true negatives + false negatives + true negatives
If each person = 1 profile photo, then to classify a photo of someone there would be 1 true positive, 174781-1 true negatives, and unknown number of false negatives.
0.9963 = (1 + (174781-1)) / 1 + (174781-1) + x + (174781-1)
0.9963 = 174781 / 349561 + x
0.9963349561(1/174781) = x
1.9925943 = x
There would be ~2 false negatives for every true positive.
So there would be a <1/3 chance the photographed person is not on VK at all, and when one searched for them when they are in fact on VK, they would only show up <1/3 of the time (the other two-thirds their subway photograph would not 'match' when tested in the dataset and it would be a false negative, and it gets much worse if the face detection has one of the lower state of the art accuracies on the other datasets). So on net you would expect to turn up a person <07 * 0.3 = <0.21 of the time. (It gets worse if we take into account the false positives as well but that particular paper doesn't report precision.)
There is no mention of any attempt to validate the findings, so I think for many of his results, he may have merely found a similar-looking person (possibly in some cases, look-alikes http://www.nytimes.com/2014/08/26/science/looking-at-twin-personality-through-look-alikes.html http://www.nytimes.com/2014/08/26/science/looking-at-twin-pe... https://en.wikipedia.org/wiki/Look-alike https://en.wikipedia.org/wiki/Look-alike ). Look-alikes are far from identical twins, so it's not necessarily surprising that
> Often, there was a striking difference between a person’s real look and the image they projected on their social network profile: a shy and grim young man might appear to be the life and soul of the party and a lover of extreme sports on his VKontakte page.
This could be due to care in social presentation. Or it could just be that they're not the same person in the first place. (The photos are not that persuasive. I'm sure there's more than one grim young men with widow's peaks in St Petersburg, or young women with curly black hair.)
- gambler 11y agoI appreciate your skepticism. And I mean that in the most positive way possible. Seems like most of HN visitors areskeptical about almost everything except extraordinary claims regarding AI prowess. But really, you're being way too generous. 99.63% accuracy was reported on a relatively constrained dataset[1] when being asked a yes-no question: "are these two faces the same"? In our case, however, we're speaking about a search algorithm that tries to match something in a huge database. The accuracy will be way, way worse. http://vis-www.cs.umass.edu/lfw/ http://vis-www.cs.umass.edu/lfw/
- Anderkent 11y agoYep, and many of the photos presented on the page don't even look like the same person.