Abby Ohlheiser reports in the Washington Post:
@deepselfie has been trained to judge (and often, insult) the composition of the ubiquitous images. Deep Selfie has its own taste in selfies, including a gender preference.
Is your selfie any good? Probably not, according to a new bot, @deepselfie, that has been trained to judge (and often, insult) the composition of the ubiquitous images.
The bot, which was created by Andrej Karpathy, makes use of Convolutional Neural Networks, a very powerful visual recognition system that, Karpathy explains, is often used for important things like recognizing anomalies in medical photographs, or looking for forests in aerial images.
“But once in a while these powerful visual recognition models can also be warped for distraction, fun and amusement,” Karpathy wrote. He taught the model to recognize good and bad selfies from a pool of two million (!) images taken from the Internet, before setting the whole thing live in the form of a Twitter bot. All you have to do? Tweet an image at @deepselfie. It will reply with its analysis.
[No, selfies have not killed more people than sharks. That’s ridiculous.]
It’s easy to imagine Deep Selfie as akin to the fictional machine it was, perhaps, named after: Deep Thought, a massive, complicated machine set on a singular task and infused with a healthy amount of absurdity. That might explain 1) the length of Karpathy’s post explaining how Deep Selfie works, and 2) why the bot keeps breaking down temporarily. Judging your selfies is hard work.
As it turns out, Deep Selfie has its own taste in selfies, including a gender preference. At the end of its training, Karpathy looked at the top 100 rated selfies from the experiment. They were all women. Here’s what else he found out, from his post:
-Be female. Women are consistently ranked higher than men. In particular, notice that there is not a single guy in the top 100.Hmm. In any case, we decided to test Deep Selfie by showing it some of the great self-portraits from art history, but the bot needed a little break before we could get very far, so we only did a couple.
-Face should occupy about 1/3 of the image. Notice that the position and pose of the face is quite consistent among the top images. The face always occupies about 1/3 of the image, is slightly tilted, and is positioned in the center and at the top. Which also brings me to:
-Cut off your forehead. What’s up with that? It looks like a popular strategy, at least for women.
-Show your long hair. Notice the frequent prominence of long strands of hair running down the shoulders.
-Oversaturate the face. Notice the frequent occurrence of over-saturated lighting, which often makes the face look much more uniform and faded out. Related to that,
-Put a filter on it. Black and White photos seem to do quite well, and most of the top images seem to contain some kind of a filter that fades out the image and decreases the contrast.
-Add a border. You will notice a frequent appearance of horizontal/vertical white borders.
Deep Selfie wasn’t very impressed by our first selection, though. “It is what it is” said the Selfie Bot to Van Gogh’s self portrait. Van Gogh gets a a score of 45.6%
It also didn’t particularly enjoy Courbet:
Luckily, we weren’t the only people who refused to abide by Deep Selfie’s singular purpose: to judge selfies. Deep Selfie didn’t like this wonderful dog, for instance:
“Your selfie gets a score of 36.4%. Not very impressed,” the bot said to the dog.
One user also discovered that the bot wasn’t as into a professional photograph of Emma Watson as it was of this drawing:
Oh, you want to see it at work with an actual selfie? Well… we sent Deep Selfie the Pope Francis selfie at the top of this post, but the bot did not respond by press time.
0 comments:
Post a Comment