Apparently, young people in Britain spend on average 48 minutes a day taking selfies.
Apparently, young people in Britain spend on average 48 minutes a day taking selfies.
But how do we measure progress in our quests to be better looking and more interesting than we are in real life
Luckily, Stanford PhD candidate Andrej Karpathy has stepped up to the plate. Karpathy has trained an artificial intelligence system to determine what counts as a “good” or “bad” selfie, based on a dataset of more than two million pictures.
The “140-million-parameter state-of-the-art Convolutional Neural Network” or ConvNets is used to recognise people and objects in photographs by identifying patterns in a lot of modern tech.
Karpathy trained the system to look at how many “likes” and “favourites” selfies on social media networks received, controlled for followers and then showed it 50,000 new pictures, asking it to decide which ones people would like.
ConvNets found that we overwhelmingly prefer selfies featuring women with long hair and small foreheads. Your face should occupy about a third of the image and weirdly, cutting off a bit of the top of your head also made it rank higher.
Not a single male selfie made it to the top 100 ConvNets rated — but the top rated male pics overall were more varied. Poses and hairstyles were much more diverse than in the top female shots.
Filters, oversaturated faces and borders also make a selfie more likeable apparently, but bad lighting, more than one person in the shot and selfies that were too close-up did poorly.
Karpathy also ran the programme on celebrity selfies, which bucked some of the trends for us mere mortals.
The famous Oscars 2014 group selfie ranked very highly despite having so many people in it and Chris Pratt, Justin Bieber, Stephen Colbert and Jimmy Fallon also managed to break the “no guys” rule.
Most of all, Karpathy wrote on his blog — style matters. Several objectively good-looking people like Taylor Swift and JLo ended up near the bottom of the list because their selfies were badly proportioned or taken in low lighting.
If you want to put your own selfie to the test, tweet it to ConvNet’s Twitter bot, and it’ll instantly analyse it.
