Land of a billion faces

A false positive in a live setting is more of an issue than it is in an after the fact setting. Because if you're getting an alert and you're running down to find the person, you have maybe a lot less time to see if it's correct. Whereas if you're behind a desk doing an investigation, you have all the time in the world to make sure you're doing the right thing. 

So a neural network is a newer form of artificial intelligence where instead of hard coding factors like the distance between your eyes or the distance between the eyes and the nose, it just learns from a ton of different examples. And what we do is we collect like a thousand examples of George Clooney or a thousand examples of Brad Pitt and the machine over time it learns the difference between those two faces and then it can apply it to a face that hasn't been seen before.

Something everyone we spoke to for this series agreed on is that these systems work best when the lighting is good and cameras are placed at face height. 

But with security cameras, that’s rare. There’s also the challenge of scale.

How do you search billions of faces or vectors in under a second? Typical databases lookup by name and email. This looks up by similarity. And doing that at scale is hard. We had to get our own data center as well for that. Typically if you're buying a facial recognition system, there's a cold start problem. What photos do you put in there? So police departments might have their own mugshots but they don't have mugshots from other police departments. And we just realized there's trillions and trillions of web pages on the internet and on social media and you know, news sites, mugshot websites.

LinkedIn's a billion dollar company, or trillion dollar company Microsoft, they don't have the right to block other people's access to public data. So it's a thing that is just kind of in an interesting spot because it is only searching publicly available information and things that people want to be private. We do know that we don't want to plaster them all over the internet. So I think that we have an instinct of what we want to keep private and what we want to keep public and that will always be the case.

We're still not even 1% of what's out there when you run the numbers, it's kind of crazy how much information is out there. So when it comes to privacy, we kind of have to look at ourselves and say, well, we are voluntarily sharing a lot of this information. Google has your location on Android all the time. Facebook knows all your habits and what you like and what you don't like. The Instagram explore tab is just like phenomenal at finding out what you like, it's kind of scary, but they kind of know a lot more information than what we do.

So already in a lot of law enforcement agencies that use facial recognition, they have a procedure saying you cannot just arrest someone based on a facial recognition match. You still have to do follow up research. So there's always a human in the loop that checks is this person, the right person? Do they have the right name? Does that person live in the same area where the crime was committed?

But what about folks who are falsely accused? 

Ton-That would argue that’s a human failing. In the same way we’re still responsible for how we drive while using GPS—when the navigation says “turn right” and it’s not safe to do so—it’s up to us and our human brains to ignore it.

And he reminds us that people get it wrong too.

He says the company did its own audit, assembling an independent review board that took a similar approach to when the ACLU tested Amazon’s facial recognition system by running photos of the US Congress past a mugshot database—which are photos of people who’ve been arrested for a crime.

Hoan Ton-That : And so they ran this independent study. But instead of searching a gallery of 25,000 mugshots it searched a gallery of 2.8 billion photos at the time. And each of the results that came up were the correct person and they went through them individually.