The image above isn't a photo of a real scene. It isn't the work of an artist either, not a human artist anyway.
This image is a fictional scene created by an experimental computer algorithm developed by researchers Qifeng Chen and Vladlen Koltun at Standford University and Intel. The algorithm is based on a class of deep neural networks known as Cascaded Refinement Networks, which is nerd-speak for: computers endowed with imagination.
Here's how it works: You give the machine a very large number of street-view images to "train" on, training here being an inhumanly tedious task of mathematical crunching. Once the training is done, out comes a computer "model" that is capable of hallucinating convincing images of street-views that never existed, similar to the ones it was trained on but never quite the same.
Let's take a moment to consider what this really is: A piece of computer software is capable of imaging fictitious scenes based on its own memories of real scenes it had seen before. If you're thinking this sounds freakishly similar to how your own mind's power of imagination works, you're not alone.
This work is a notable yet very recent improvement upon already existing systems known as Generative Adversarial Networks (GANs), which have been applied successfully to many real life tasks. GANs have been used to recover partially damaged images, colorize old black and white photos, as well as improve low-resolution images, all by hallucinating missing image details to match what's already there. They have also been used to make up digital scenes for computer games, or even dream up entirely fictional images based on textual descriptions only.
That's what is already out there today.
Extend this line forward through time, and you'll find that very interesting, if disruptive and controversial advancements are going to start happening soon. Advancements that are going to challenge our current primitive notion of intellectual property in ways that were unimaginable before.
In the next five to ten years, I claim to you that many of following AI content-creation advancements will become available:
Computer generated instrumental music tracks
The easiest would be electronic music like new-age and trance. Picture a machine capable of analyzing all the works of Darude, and generating convincingly similar tracks in bulk for mass consumption. The concept is very similar to image generation with GANs. In fact, deep learning is already being used to synthesize speech in devices like Siri and Alexa. Synthesizing instrumental music is a hop away from there.But, what would Darude think about that??
Computer generated paintings
Trained to imitate the style of Picasso or Monet, a deep neural network can be made to create interesting art at an unprecedented scale. You could even "commission" a computer artist, by providing a description of what you want the art to look like, and a sample set of paintings that you happen to like, and voila, your personal dedicated AI artist can spew hundreds of thousands of candidate art pieces for you to choose from.Here's an example of a synthetic painting created by supplying a real photo to an AI that was trained to imitate the style of Van Gogh [source]:
If you're an artist mocking those factory workers who keep complaining robots are taking their jobs, don't feel so safe.
Computer generated pop songs
My favorite pop singer (Amr Diab) is getting old. Luckily his repertoire includes about 100 solid songs and his style is consistently awesome. I have always been mulling the idea of building an AI machine trained on all his songs to synthesize more songs in his style, forever. When I say in his style, I mean lyrics, instruments, beat, sounds effects, even his voice and intonation. Every aspect of what makes a song a song can be convincingly imitated by a well trained AI machine, in theory.In practice, I think we are a few years away from that, but when we get there, I wonder what would Amr Diab himself think about this?
There are laws protecting a singer's own songs as intellectual property. But what about songs synthesized by a machine trained to imitate the singer's style? Who owns the intellectual rights for those songs? And what if I were to "compete" with the real singer by flooding the market with hundreds of songs that sound like they're his, but available for a fraction of the price?
Computer generated novels
Further down the road now and much harder to generate, but what if I could provide you with an unlimited supply of synthetic Dan Brown thrillers or synthetic Danielle Steele love stories, whatever your taste desires? Would a novelist have any control on using his/her own novels as training material for a novel-writing AI? Who would bear the burden of proof in a case involving an accusation that a synthetic novel was in fact based upon the style of a real novelist? How would you even go about proving such an accusation?Beyond images, sound, and text
What about computer-generated movies? What about computer generated schematic designs? How about computer generated building architectures? Interior designs? Landscapes? City zoning plans?A lot of questions, and I don't claim to have any of the answers. But this I know for sure: our concept of intellectual property is still in its infancy, and it's going to have to grow up in a hurry.
Comments