If there’s one thing the internet needs it’s more cat pictures, so researchers from Nvidia and Cornell University developed an algorithm that will turn pictures of dogs into pictures of cats.
This neural network—a type of computing architecture loosely modeled on the human brain—was developed by a few of the same researchers behind the algorithm that can turn winter into summer in any video and employs similar principles.
As detailed in a paper published to arXiv, the neural net is actually a generative adversarial network (GAN), which is a way of training a machine learning algorithm without human supervision. In GANs, two neural nets are pitted against one another: One neural net generates new images and tries to trick the other neural net into thinking the images are real. If the other neural net is able to tell the generated images are false, the neural net generating the images keeps tweaking its algorithm until it generates images that are nearly indistinguishable from the real deal.
In this case, the neural nets were first fed images of dogs. A neural net would take the dog image, break it down into code describing various features of the dog photo and then combine these features with various “style codes” derived from features taken from an image of a cat. Each iteration of this mashup of dog features and cat features would result in a hybrid image, which would then be fed to a neural network tasked with determining whether the image was genuine.
The researchers also applied the technique to other inputs with similar results. They managed to turn cats into dogs and jungle cats into house cats, and to turn line drawings of things like shoes and handbags into realistic-looking products.
According to the researchers, teaching a machine how to map images of cats to dogs and vice versa was particularly challenging because there are so many different features to take into consideration. Not only does the algorithm have to account for factors like head position and fur pattern, but it also has to separate the meaningful image content (that is, the face of the animal) from the stuff that doesn’t matter, such as the background.
The researchers wrote that they now hope to apply the technique to more complicated scenarios, such as text and video. It’s kind of like a 21st century take on The Island of Doctor Moreau—HG Wells’ novel about a mad scientist who creates horrific animal-human hybrids—but there’s no secret island laboratory required. In fact, the algorithm and sample data sets are freely available online.
Photographs and video used to be the gold standard for representing reality. Yet as machines learn to distort images to create their own versions of reality and these tools become widely available, it will be harder and harder to distinguish what’s real and what’s not—for better or worse.
Motherboard’s documentary series “Dear Future” was nominated for a Webby. We’d love your vote, and it only takes a minute.