Home » Science & Technology » Selfie-Esteem: this Stanford PhD built an algorithm to rate your selfies

Selfie-Esteem: this Stanford PhD built an algorithm to rate your selfies

Asad Ali | Updated on: 13 February 2017, 7:41 IST

Turns out, just like good Taliban and bad Taliban, we now have the concept of good narcissism and bad narcissism - or as they're called these days, selfies.

Andrej Karpathy, a Stanford PhD student working on deep learning just finished a rather unique project.

He made an artificial intelligence (AI) network 'consciously' sift through millions of selfies, and select the best ones.

How does it know what best is anyway? Serious programming went into that.

The AI behind it is something called Convolutional Neural Networks (ConvNets in short). Karpathy explains in his blog: "if you're seeing or reading anything about a computer recognising things in images or videos, in 2015 it almost certainly involves a ConvNet."

Andrej Karpathy uses artificial intelligence to 'read' selfies and rank them

He was so thorough that the only real problem with this experiment is understanding why it was done at all. He could've used his technical chops to give us an Interstellar for Dummies, which is more goal-oriented and serves an actual purpose for those still nodding nervously when friends praise the film in a social setting.

Instead, he decided to spend his free time training artificial intelligence to be like a highschool cheerleader or jock obsessing over their nth photograph, selecting just the right one for upload.

The tech is pretty darn impressive

Karpathy wrote a computer script that trawled through 5 million Internet images tagged with #selfie and picked out those which showed only one face.

The software was left with 2 million such single-face images.

Next, based primarily on the number of likes (controlled for the number of followers) the selfies were labelled good or bad.

All this data was fed into the ConvNet.

It was then shown 50,000 completely new selfies which it assessed using the pre-fed data as a sort of frame of reference. The results were largely unsurprising, unless you've been living on Mars.

The (not so) Big Reveal

Female selfies are more popular. Especially those where the woman's face is "slightly tilted, and is positioned in the centre and at the top of the frame," and oh, you better have "long strands of hair running down the shoulders."

In fact in ConvNet's ranking, not a single male photograph made the top 100!

Well, duh.

If there's a penny to be made for every such selfie being taken in real time in Haus Khas Village (Delhi's has-been-hipster hood), there'd be a lot of millionaires.

Female, long hair, tilted head, filters: the anatomy of a perfect selfie according to artificial intelligence

Still, Karpathy did analyse what makes a popular male selfie. "A slightly broader shot with head fully in the picture, and shoulders visible. It also looks like many of them have a fancy hair style with slightly longer hair combed upwards."

Want to know how yours stacks up? You can tweet your selfie to ConvNet's twitter handle and it'll respond pretty quickly, showing you where your selfie stands in relation to the beautiful world out there.

You can also use ConvNet to find the perfect crop for your shot: it takes the selfie, tries various combos of crops and applies the one it thinks is best for the photo. Sometimes the actual person can get cropped off in ConvNet's best crop selection though - an inside joke sure to make Robocop chuckle.

First published: 3 November 2015, 6:50 IST
Asad Ali @asadali1989

Asad Ali is another cattle class journalist trying to cover Current affairs and Culture when he isn't busy not saving the world.