Selfies are an amazing way to tell the world "here I am, rock you like a hurricane!" But according to a new artificial intelligence (AI) system built by a Stanford University researcher, not all selfies are considered equal.
Stanford PhD student Andrej Karpathy built a deep learning system that analyzed 2 million selfies and could tell which selfies would attract the most likes.
It turns out, if you want your selfie to take over Instagram, you might want to follow a few rules — be a woman, have long hair, take a close up, and crop it close enough that the forehead is cut off, among other things.
But before the AI got to these conclusions, it had to be trained.
Here's how it works
The AI system is based on a technology called convolutional nets, which were first developed by Facebook's head of AI research Yann LeCun in the 1980s. If you've ever used image recognition or deposited a paycheck at an ATM, you've used a convolutional net.
The selfie-judging AI system works like an assembly line — the image goes in on the left, goes through levels of analysis, and comes out on the right. Each level breaks the image down pixel by pixel. The first few layers look at simple facets, like shapes and colors, while the layers toward the end look at "more complex visual patterns," Karpathy writes.
Karpathy found the selfies by looking for images tagged #selfie, then divided them into good and bad according to the number of likes (taking into account the number of followers that the person had). He also filtered out images that used too many tags, and people who had too few followers or too many followers.
Then the magic began. According to Karpathy, the system "'looked' at every one of the 2 million selfies several tens of times," and found the components that either make a selfie good or bad.
After the system was trained, he fed it 50,000 selfies that the AI had never seen before, and it was able to "rank" them based on the images alone from good to bad. In the image below, the AI system ranks the selfies from good to worse.
The AI found that the selfies most likely to get the hearted had a few things in common. They often contained long-haired women on their own. The selfies were also very washed out, filtered, bordered, cut off the forehead, and featured a face in the middle third of the photo.
Below are the cream of the crop. Notice that of the best selfies, not a single man is included, and there are very few people of color.
On the other hand, the worst images, or the selfies least likely to get any love, were group shots, badly lit and often too close up.
Get your selfie judged by a robot
Karpathy even made a Twitter bot, which looks at people's submitted selfies and judges them automatically. I tried it out myself with my latest Instagram selfie from about two weeks ago, and got a slightly better than average score.
It's also pretty fast — it replied with my results in just a few seconds. Try it out by tweeting a square image or link to an image at @deepselfie.
But it might be a bad idea to follow the rules to a tee just for the likes. The AI system is more like an amalgamation of the things that a lot of people like, excluding any sort of creativity like funny faces, blue wigs, or pictures of you and your friends.
After all, selfies are supposed to be an expression of self-love. So if your favorite selfie doesn't score that high, who cares — you just do you.