Dreams, memories and GAN
Why are generative adversarial neural networks so popular today in the art scene? To me it is partly because the output so much resembles how the unconscious serves us dream elements while we sleep, often showing only vague or incomplete imaginary, yet still detailed enough to be recognized.
And as Ivona Tau sees it, memories have the same imperfections as GAN output in the images they produce. Although the shapes are imperfect, we still know/feel what they represent. To illustrate this theory, I did many experiments in the past whereby I made sketches from details of places I had seen the day before and although, when compared, they had almost nothing to do with the reality, people did recognize what I sketched.
Here is an example of a detail in the central station of Antwerp:
And here is the real place:
Photo of that booth in station.
As you can see, there is little resemblance. Yet, when I showed the sketch to my friends, they knew exactly where is was.
I did the same test with the BatterSea Power station in London where the chimneys almost looked like bottle necks.
The strange thing about this experiment was that I really did my best to draw the scene as accurate as possible, but my memory became distorted by the interference of various other memories and confabulating (filling gaps in memory with fabrications one believes to be true). In an unscientific way this is comparable with how GAN assemble an image from various sources. Where it doesn’t find a suitable reference, it will try to recreate one, not always succeeding to come up with the right solution, though. Experimentation and patience is key when working with gan.
The idea of gan is of course perfection. But we aren’t there yet and the imperfection makes GAN so immensely interesting because you get results that you otherwise wouldn’t dream of. This will change with time and therefore this is the right moment to enjoy it the most. Because once it becomes mainstream, the fun and magic is over.