Artificial Intelligence in Art
For years and years, artificial intelligence was quite stagnant in anything remotely resembling artistic creativity.
AI could launch a missile or two while embarrassing you in a game of chess — but had some real writer’s block issues, had no great demos to torture musical producers with, and was especially terrible in visual arts.
Yes, there were some very brave attempts to break through the creativity barriers, but on average we were mostly getting shy William S. Burroughs cut-ups, poorly interpreted Van Goghs made out of dog photos and cat-on-a-keyboard music pieces.
One of the latest — and more promising attempts in the AI creativity — are based on so-called Generated Adversarial Networks (GANs). In GANs there are not one, but two neural networks — called Generator and Discriminator — who compete with each other.
The function of Discriminator is to compare the output of the Generator to a dataset of reference images of human-generated art — and punish the Generator for performing badly.
Basically, it’s an Artist + Critic cojoined Siamese twin — or a very self-conscious Artist who constantly whips its own butt for being bad.
While some of us can perfectly be very self-conscious — to a point where it’s a problem to enter the sweet void of night dreams — it is hard to compete with the referencing abilities of the Discriminator.
We can only remember that embarrassing moment in the 6th grade and our inability to come up with a smart and snappy remark last dinner — while the Discriminator can vividly recall tens of thousands of reference images in a blink …I really doubt it ever sleeps at all.
The problem here is that the Generator is more trying to fool the Discriminator than to actually invest time and effort into making an actual art — which is amazingly human — but requires quite a lot of fine-tuning to produce anything impressive.
One of the most interesting GAN-based art attempts were recently performed by Robbie Barrat:
