They're not meant to. This is the development of base, mechanical technique, not of composition and theory. Lots of artists paint fairly mediocre pieces when developing technique and very often copy other works of art in the process.
> I don't believe that neural networks will ever be artists because they work on copying existing patterns. Until there is an element of randomness fueled by context, they aren't creating anything new. That's where genius lies.
What you ask for is actually a fairly simple architecture, it's just presently computationally intractable. So we're seeing if we can optimize the approaches to run on modern(ish) hardware.
Eventually an orchestra agreed to play one of the works. People liked it. Until they found it was composed by a computer, and then they hated it. People are insanely biased against computers.
That said, the algorithm was not a neural network. NNs do have problems producing music, they forget long term patterns quickly and so the works are easily distinguishable from human music. But it's possible this could be solved in the future with some new techniques at training NNs with longer term memory. I don't think it's long before the world is full of AI produced music. And you won't be able to tell.
I suspect reframing of problems, seeing things in new light, paradigm shifts, new ontologies, whatever you want to call it are not quite so simple as context-dependent randomness! I don't think we understand this process very well right now.
Finally; I think there is an artfulness in copying existing patterns, because the way in which you "abduce"[1] an observation-explaining theory out of the infinite space of possibilities is a creative and aesthetic process.
Lil Wayne has a wonderful line in his song 6'7', "real G's move in silence like lasagna". Unbelievable.
Let's extrapolate that pattern:
"Real H's move in silence like phonebooks".
Not really the same ring to it, huh? It's not just pulling out a silent letter to emphasize the silence, but there's the context of Lil Wayne being a rapper and self-proclaimed G (gangster.)
Is our extrapolation "new"? Sure, in that no one has (likely) ever said that. But while it mimics Weezy's style, it doesn't understand it's context. Similarly, if Katy Perry sang the same line as Lil Wayne, the context doesn't make sense (Katy Perry is no gangster...)
EDIT
A better example, specific to art, would be Warhol. He explicitly copied real-world objects as art, to create something new. But the newness wasn't that he made a clear copy, it was that his copies reflected the shift in materialism that came with mass-production. Warhol mass-producing "art" was a social comment that resonated _at that time_.
His art was more than the process, it was the context in which it was made and what that said more broadly about society. That's why it resonated.
Sorry, I'm too thick. I mentally thought of a gangster moving laterally, splayed out and wondered how it would be quiet. I had to google this to understand that he meant silence like the letter 'G' in 'lasagna'.
The end result here is that in English the g causes a sound change after the n, and basically qualifies it as a silent letter. This is a subtly different sound from the Italian version, where it merges with the n and does not qualify as a silent letter.
[0] https://blog.acolyer.org/2016/04/21/the-amazing-power-of-wor...
[1] https://www.theguardian.com/technology/2016/may/17/googles-a...
Dependent randomness? or, in other words, deterministic randomness? a paradoxical notion
That said, when people express skepticism about current AI produced art, what they are really trying to express is that current AI lacks a theory of mind in its productions. Artists are not simply sampling from a generative model, they're generating a construction informed by complex inner states while also predicting how arrangements of states will affect a perceiver. Great artists, especially writers and poets, are able to operate at above average levels of intentionality.
Until we have AI with basic commonsense, the good news is—for an undefined period of time—AI will foreseeably be tools which augment artists (humans) and raise the floor for what counts as creative work. Even within a couple years, things by talented teams that would impress today will be run of the mill thanks to intelligent tools but one can hardly imagine what talent of the future enabled by semi-autonomous tools will create. This help could hardly be more timely thanks to rising production costs, falling profits and new content intensive modalities as VR.
If you're interested in theory of mind, this is a good recent read: [The Complexity of Jokes Is Limited by Cognitive Constraints on Mentalizing](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4864470/)
Art comes from life. AIs won't experience life until they have bodies. Until then the best they can be is a psychotic imitation machine chained to a human-selected data point.
That can have quite a lot of value, when paired with humans who can select data points intelligently. But we still don't even have a whiff of an understanding of how to create autonomous intelligence.
...BUT I would also say that this technology is yet-another-one which can act as an effective 'force multiplier' in the sense that if you already _have_ a reasonable 'eye' you can use this as a very very powerful tool.
One of the benefits of that is that if like me you have a 'decent' eye only, but no e.g. visual genius, an app like Prisma plus a modest amount of straightforward judicious additional editing / cropping, can lead unnervingly quickly to results which I find _very_ satisfying.
To put that another way, I find that I can now in-phone produce images which for my own interests and tastes are quite arresting.
They are not fine art, or innovative... but they represent a rendering into pixels of notions I am not otherwise able to execute, and frankly am not able to even foresee as effective.
I find then that I am often in a very specific aesthetic mode using them: discrimination and rejection of all but the strongest candidates.
Feels like I am executing a GA selection function, essentially, and then just doing a little post production...
We don't have to have "deeper meaning". We just have to pass enough double-blind trials of whether an expert can distinguish "deep meaning" from "AI-generated meaning". I'm not a big fan of the "humans are special and machines can't capture magical human nuances" hypothesis.
We might not get world-changing, paradigm-changing, transcendental, strong social commentary-inflected art, but we are getting some cool stuff people like looking at. Not to mention meme material ("Dog or Bagel" and friends) that uses stuff like this as part of its input (which, I would argue, is its own kind of art.)
A neural net has plenty of randomness (which shouldn't be confused with nondeterminism) inside it, right from the start with its randomly initialized weights & biases; no two trained style networks will be exactly alike and produce exactly the same output unless one takes major trouble to do so.
Any artist (or writer) has had the experience of people finding meaning and depth in their creations which they had no idea about, and may have thought one of their worser works. This can reflect unconscious work, of course, but it also reflects that the artistic process is as much about throwing random things at the wall to see what sticks and triggers an intuitive reaction in other people that 'hey that's good!'. One may have fail hundreds of times before getting a hit. Rarely does one sit down and think 'I am going to write an intricate story about dying in Venice with such and such symbolism, a buildup to the ending, and referencing these specific past authors', write it out in full in 1 pass without revision, decide it's good, and have everyone else agree. Art is as much about about editing, curation, and recognition as it is about planning. A neural net could create great works of art without any planning or intention of doing so, if the right person sees it and selects it out of the vast stream of images the net will generate.
Partly because they are like 200x200 pixel images. There is no power in the original either that way. And I've been seen a ton of these posts like "look at these beautiful images", and the images are absolutely tiny. It's really ridiculous and annoying.
Not to mention the transformations the artists made that the software cannot: turning pinpoint stars in the sky into huge pinwheels and fireballs; turning a human face and hands into an inhuman creature.
Capturing the semantic meaning of the distortion would another increase in complexity, but perhaps not necessary, as the software doesn't have to understand why we selected "The Scream" as the style we want (although it would certainly be cool to be able to specify "Use the style of "The Scream", but expressing melancholy instead of horror.").
> Notice how the transferred images have none of the power of the originals.
I disagree. Maybe that's because I've been described at emotionless but the original art doesn't have any "power" to me at all. I honestly see all of them as basically the same as far as impact goes.
Now I'm likely in the minority here but it shows that art is going to be very personal and there isn't going to be very objective ways to measuring "power" or "emotional impact" except through an aggregate polling.
https://www.instagram.com/p/BL-ngYTjXZH/
https://www.instagram.com/p/BLfxvgyD-mu/
https://www.instagram.com/p/BHnZyddjumg/
This is a very new field, I find it very exciting and can't wait to see where we are in 5-10 years, personally.
That said, I think saying never is a bit strong. People like an amazing variety of things, and given enough design, statistical modeling, luck, and randomness something interesting enough for a listener to infer meaning can pop out. You should check out this article about David Cope's work [0] and thoughts on creativity.
[0] https://www.theguardian.com/technology/2010/jul/11/david-cop...
Judging from some of the stuff on display in galleries right now, I'd be shocked if machines can't outperform or match. And they'll only get better, right?
That argument could be said about any innovation.
The Google paper is an iteration on neural style transfer; just because it doesn't mimic the "soul" of art doesn't mean the field is not worth pursuing.
This is effectively the 2d image version of those "random text generators" that ape a given style. It's really interesting and cool but it's not "Hollywood AI" in any sense.
Take the first set of images, in the style of “Head of a Clown”, by Georges Rouault. Not only does it not recreate the brush-strokes, but the content of the restyled image doesn't fit the artist's style.
In the case of Munch's The Scream, and Starry Night, and every Monet, the subject matter is as important as the image. Maybe you'll be able to mask your Facebook profile in a combination of Murakami and Pissaro, but that's not the extent of their genius.
Similar things are attempted in music. While we can pretty programmatically formulate pop music, deeper meaning eludes machines. Bob Dylan didn't just write great lyrics, he tapped into the zeitgeist of the era in which he wrote. When Kendrick Lamar or Beyonce write about black struggle in America, that's in the context of a larger societal conversation.
I don't believe that neural networks will ever be artists because they work on copying existing patterns. Until there is an element of randomness fueled by context, they aren't creating anything new. That's where genius lies.