It's been a week of nightmarish image-making, thanks to the public release of Google's "DeepDream" algorithm. If you missed this whole thing, let me explain: DeepDream is an artificial neural network that Google originally developed to identify the contents of an image using computer vision. To do this, the network is "taught" using thousands of existing images -- say, of animals -- which means it's possible to reverse the algorithm to actually create images based on what it already "knows".
Long story short, this artificial intelligence -- originally built to help computers "see" images -- has been turned on its head and used to create new images instead. The internet has put it to work this week, running all kinds of pictures and movies through it. But today we heard from an artist who might be the first to actually incorporate Google's AI into his work.
C. M. Kosemen is a Turkish painter who describes himself as a surrealist painter specialising in amazing, uncanny animal forms -- his paintings are, in a way, already similar to what DeepDream churns out. Kosemen noticed the similarity, and is now applying DeepDream to enhance and further his own paintings. The results, as you might expect, are horrifying and amazing.
Kosemen first learned of DeepDream a few weeks ago, when Google posted about the project, and immediately saw a relationship between its creations and his own work. "I was also a bit jealous -- the DeepDream algorithm's results were like a turbocharged version of my work," he says in a statement. "Here was a machine, imitating something very much like an artist's instinct."
So he enlisted the help of a Swedish AI researcher named Roelof Pieters, who studies deep learning, to put Google's new GitHub code to work. He ran a series of Kosemen's paintings -- which feature birds, monsters, demons, and all manner of unworldly half-animals -- through the algorithm, and found that it was very good at enhancing his work. Like, shockingly so.
In one painting, a lizard with a demonic head crawls along a crag of grey paint. DeepDream turned the lizard's leg into the head of a deer, then shaped the face into a gorilla's -- the fog in the background? Now, it's a thick wall of worms, lizards, and other ambiguous reptiles:
In another painting, DeepDream created a series of "gimbies", which is Kosemen's neologism for the little creatures the algorithm creates. "One of my paintings showed a bird-headed demon with open hands," he writes. "I looked at the processed version made by the AI, and there was a 'gimby,' sitting comfortably in its hand:"
Even stranger? Kosemen and Pieters point out that the creatures DeepDream creates recall demons from historic sources -- as though, by improvising using its knowledge of animal anatomy, the algorithm had reverse-engineered the things humans fear. He compares one to a medieval painting of a demon:
And another to a Gnostic demon called Abraxas:
Or even this multi-limbed monster, also medieval:
It's not that Kosemen is saying Google's AI can dream up demons. He's saying that it uses the exact same sources that our brains do -- the world around us, with all its natural forms and shapes -- and ends up producing imaginary beasts that aren't too dissimilar from our own imaginary fears. Here's his fascinating quote:
I believe that certain forms of animals - snakes, wolves, birds, insects, are hardwired into our brains. I also think that this reflex is directly responsible for the demons, angels and monsters in many religions and mystical experiences today. What the DeepDream algorithm did to my work reflects this transformation - the evolution of mystical imagery in the human subconscious. There are reptilian stares, celestial birds and strange beasts with a thousand feet, creeping in and out of every corner. It's like watching the birth of a new subconscious - a digital one.
Of all the DeepDream projects we've seen over the past few weeks, this is by far the most interesting. Not because it's art -- but because it shows how DeepDream learns and imagines in a way that's not so very different from our own brains.
Check out a few of the duo's works below. Or is it a trio?