Horny Nerds Used Artificial Intelligence to Stop Seeing Tattooed Women on Tinder

Professor Frink was not involved.

|
Sep 2 2015, 7:08pm

Image: Flickr/Alley Oop

If you're anything like me, Tinder's romantic dystopia is mostly an exercise in denial. As photo after photo flips by—is that another person grinning while draping their lily white arm over a sad, drugged-up tiger?—the whole thing starts to come off like a Nascar race; "left" is the only way to go, and fast.

Why doesn't Tinder just show me people that an algorithm has already determined will be appealing to me?

I solved this problem by deleting the damn thing and re-entering the world of physical human relationships (it's not so bad, trust). But Harm de Vries and Jason Yosinki, computer science PhD students at the University of Montreal and Cornell, respectively, took a slightly different approach: they built a complex deep learning algorithm that can recognize the type of women they like on Tinder.

"I moved to Montreal to start my PhD here and started using Tinder," de Vries told me over the phone. "I consistently didn't like a type of girl. In my case, a lot of tattoos and piercings and stuff. It kept showing me these pictures while I was completely disliking them. I was like, yeah, they're definitely not using my past history of the types of girls I like."

The algorithm is described in a paper published to the ArXiv preprint server. It has not been peer-reviewed.

"It's just a matter of time before someone commercializes this"

The deep learning approach the researchers used is called "transfer learning," in which a neural net is trained to recognize one thing (in this case, gender) and then some of those learned parameters are used to recognize another (attractiveness).

The network was first trained to recognize gender using nearly 50,000 OKCupid photos, and then trained on attractiveness with around 10,000 Tinder photos. The final setup achieved 63 percent accuracy in recognizing people de Vries might find attractive on Tinder from their profile photos. When the images were fed through a pre-trained network, not optimized for gender or attractiveness, the accuracy rating peaked at 68 percent.

"It's just a matter of time before someone commercializes this," de Vries said. "If you have a lot of users swiping, you could train the neural net directly on those, and it would be much better than what I have now. This is just a first step."

Okay, so the result wasn't perfect. In fact, this extremely complicated approach is probably only marginally more effective than this robot pen that endlessly swipes right because nothing really matters. I guess de Vries will just have to deal with all those *shudder* girls with tattoos.