Editor’s Note: This article was originally published in an edition of NOVA’s email newsletter, NOVA Lens, and has now been repurposed for NOVA Next. Sign up for NOVA Lens here (select “NOVA Newsletters”).
Mine your old Facebook profile pics, find a snapshot from 2009, and proudly show off how much you’ve grown to your family and friends.
Seems fun and positive, right?
This viral “#10YearChallenge” might seem like a harmless meme, but rumors are spreading that its intent might not be so benevolent.
Last Saturday, Wired columnist Kate O’Neill published a piece (and tweeted the message below) suggesting that Facebook could be using data from the 10YearChallenge to further train its already advanced facial recognition algorithms. Some of the subsequent buzz over the Wired piece has gone so far as to insinuate that Facebook may have planted the seed for the meme in the first place—a theory that, while unproven, isn’t necessarily outlandish.
“Facebook probably has one of the strongest facial recognition algorithms,” says Anil Jain, a computer scientist at Michigan State University. That’s in large part because more than one billion people are active on Facebook—which gives the company a huge leg up compared to its competitors (like Amazon, Microsoft, Google, and IBM).
But despite having such a gigantic data set to work with, Facebook might not have facial information that is tagged by age—in other words, the date a photo is uploaded to Facebook doesn’t necessarily correspond to the moment in time it captures. When participating in the 10YearChallenge, users explicitly state that their two photos are 10 years apart.
So unless a significant percentage of the participants are lying, the 10YearChallenge clearly fills a need. “Facebook could use this data to improve their algorithm’s capability to do age-invariant face recognition,” Jain says. Age-invariant face recognition refers to an algorithm’s ability to recognize a person’s face many years after the system first saw that face. It can also refer to the algorithm’s ability to identify two photos of the same person taken as much as 10 years apart.
Beyond the sheer availability of tagged (or labeled) data, facial recognition across time remains difficult. As the age gap between two images increases, a facial recognition system’s prowess begins to break down. To study this phenomenon, Jain and his colleague, Lacey Best-Rowden, acquired mugshots of some 18,000 repeat criminal offenders from the Michigan State Police. Each subject had an average of eight mugshots in the database over an average span of 8.5 years. Jain’s and Best-Rowden’s study showed that state-of-the-art face recognition starts to degrade after a time gap of seven or eight years. “I think that probably accounts for the fact that they [Facebook] may have internally evaluated that their facial recognition system is falling after 10 years or so.”
Erik Learned-Miller, a professor of computer science at the University of Massachusetts, Amherst, is a bit more skeptical of the idea that Facebook might be the mastermind behind the meme.
“It’s certainly possible that they could have sat around scheming about this, but I kind of doubt that,” he says. “It’s pretty hard to predict what’s going to go viral and what’s not. That being said, they do have access to enormous amount of data, and now this is just another piece of data that would be quite valuable in creating something to understand age progression.”
That would mean Facebook could hold the key to developing a model for age growth, or it could figure out how to target advertisements to people based on whether or not they have wrinkles, for example. And these developments would set a precedent for decades to come.
“In the very near future, [we may have] electronic advertisement screens which, as you are walking by, will notice that you’re a female in a certain age group, and that you’re Hispanic, and so on,” Jain says.
Refining age estimation technologies wouldn’t necessarily be a totally commercial endeavor, though. Some are working in the field for social good. For example, Jared Rondeau, a PhD student of Learned-Miller’s, is using facial analysis to perform age estimations in order to predict whether or not a given image online contains child pornography. If it does, the application (called DeepPatrol) flags the image file as illicit material for review by law enforcement.
“In the future, I hope to incorporate facial recognition to check databases of missing children to try and identify potential victims of human trafficking,” Rondeau said in an email.
Learned-Miller says that these social uses of algorithms are important to pay attention to right now—as O’Neill writes in Wired, what Facebook may or may not be doing with the 10YearChallenge data is something of an inevitability.
“I was quite pleased with [O’Neill’s] article,” he says. “I think it struck a good balance. For people who are freaking out about it, I would encourage them to put their attention in a slightly different place. I’m more worried about things like: How good is the face recognition that police departments are adopting, and does it work well for different subpopulations? I think those are bigger and more important issues right now.
“I don’t blame anybody for being a little bit paranoid, but I think there are better places people can put their efforts if they actually care about where we’re going to be in 20 years.”
Still, Jain warns, "We are providing a lot of digital data about ourselves and we don’t think twice about it.”