This Artificial Neural Network Will Tell You If Your Selfies Suck

This Artificial Neural Network Will Tell You If Your Selfies Suck

Ever wondered how your own selfies compare against the internet's vast amount of artistically crafted narcissism-laden self portraits? Turns out, there's an AI for that.

Feeding two million internet selfies into a "140-million-parameter state-of-the-art Convolutional Neural Network," a Standford PhD candidate has trained an AI to classify "good" selfies from "bad" ones. Convolutional Neural Networks, ConvNets for short, are often used to recognise objects, places and people in photos by following common patterns.

Andrej Karpathy, who's worked with Google Research and DeepMind, notes in his blog post that "if you're seeing or reading anything about a computer recognising things in images or videos, in 2015 it almost certainly involves a ConvNet."

Filtering five million internet images tagged "#selfie" down to images containing only one face, the software was left with about two million images. The neural network was trained to evaluate whether a selfie was "good" or not by looking at how many likes these selfies had on social media (the blog post notes that they controlled for the number of followers). Karpathy's program then analysed these images by breaking down each one into layers of shapes and colours.

Then, taking 50,000 new selfies — ones that ConvNet hadn't seen before — they watched as the program perused the images before making a decision about which ones people would like.

It found that we prefer female selfies. Specifically females with long hair, slightly tilting their head while the top edge of the frame cuts off their forehead. A smaller forehead and longer hair may be more desirable because it reminds us of younger women. And as expected, we also like filters, over-saturated faces, and borders.

Men got the short end of the stick here. In a list ConvNet made, men didn't even make the top 100 "good" selfies cut. The good news is that men don't have as many selfie standards to follow. ConvNet found that the most popular mens photos showed them having their full head and shoulders in the shot. Sporting a fancy hairstyle that's slightly longer and combed upwards did well, too.

ConvNet does have some selfie-dont's that include taking pictures in low lighting, in large groups, or with your head framed too large.

This Artificial Neural Network Will Tell You If Your Selfies Suck

Continuum of the best (top) to worst (bottom) selfies, as judged by the ConvNet. Original Image Via: Andrej Karpathy blog

If you're curious to how your selfie stands up, tweet it to the network's Twitter bot. It claims to have a 60 per cent accuracy rate while giving you results in under a minute.

Image via: Andrej Karpathy blog

[Andrej Karpathy blog]


Comments

    They all suck?

    The AI they used might be great but the design of their method was purely moronic. Measuring using "likes"? Are these people sheltered innocent babies that have never used the internet before?
    AIs like this are only as good as the data you feed into them and in this case you feed it shit, you get shit.

    Maybe if they'd have used the mechanical Turk system to get people to evaluate the images they'd have gotten better data to base the system on.
    "Likes" aren't given to show the quality of anything, they're more about showing support, boosting someone's ego, or trying to gain favour -especially when it comes to males "liking" the photos of females.

    Last edited 30/10/15 2:43 pm

Join the discussion!

Trending Stories Right Now