Seems to me this is one case where AI has not surpassed humans:

https://www.bbc.co.uk/news/technology-45980863

But, if I'm right, why did it flop in art and flip in go?

Not surpassed humans... yet.

Two things.

First, it did not try to surpass humans, but to imitate them.

Second, it's art. Didn't a painting produced by dipping a mule's tail in paint sell for something like the equivalent amount of money (adjusted for inflation) early in the 20th century?

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.

Depends on your measure of success, I suppose. If I could paint something and sell it for over $400,000, I'd consider myself a (financially) successful artist.

The output of a program all depends on what you ask of it. If we just wanted a realistic looking portrait, it certainly wouldn't be hard to make a computer that can take a real life scene and apply a transformation that makes it look hand-painted. There's almost certainly an app that can do that available for my phone.

I think the piece of art did exactly what it was supposed to do. Whether that was aesthetically pleasing or a worthwhile artistic statement is another question all together.

AI flop?

No, human art market flop!




Thanks captain obvious.

While I appreciate the work they put into the software, the price tag seems a bit too much, to me.

It's just another case of humans trying to pass clownery as art for profit. Nothing new here.

And to be clear: the technique used to generate the portrait (no details provided in the article, but most likely GAN) is not clownery, just the process of selling it as a piece of art.

I believe this is appropriate here:

https://xkcd.com/1263/

It was indeed generated by a GAN (Generative Adversarial Network). Some of the backstory is actually pretty interesting, if you are interested in that sort of thing (if you're not, just ignore!): https://www.theverge.com/2018/10/23/180 ... arrat-gans

I deliberately didn't try to steer the discussion - pointless anyway here because someone always derails the thread. But I had hoped the deeper point that I saw in this development might emerge. It didn't.

What I had latched on to was that the "art" work produced was (to me) absolutely awful. The price paid and whether or not it counts as art were immaterial. The point is that it was awful and much, much worse than a good human painter could do. (You can take the cargo cult approach and say the AI picture is redefining art, or similar - but that's all bullshit.)

The argument that it was imitating humans and not trying to surpass them is also off beam, I think. AlphaGo (pre-Zero) set out to imitate humans and still surpassed them. I'm assuming that the art AI researchers at least imagined the possibility that their program would have achieved the same quantum leap effect as AlphaGo, and that now they are just making the best of a bad job.

So, on the surface to me as a non-expert, we have two programs trying to replicate something humans have so far done rather well (painting and go). In one case the result is awful. In the other it is extremely good. Why the difference?

That is what I was hoping to find out. My own speculation is that painting and go are just two different - and too different - activities, and go happens to be the one humans are not as good a. The source of the difference (I continue to speculate) is that go involves a lot of calculation, and we are bad at that; computers are good at it. Painting involves some other attributes. I'm not sure what but I suspect it lies in the way we communicate with other humans. Computers are bad at that. In go we are not trying to communicate. Painting seems pointless without communication.

To put all that another way, I suspect the art AI has demonstrated to us that it is indeed calculation that makes go AI bots superior: not fancy strategies or new josekis. I don't think go AI has come up with any new go wisdom to teach us, and maybe never will. The best it will do is to put probabilities on certain proverbs and heuristics working, helping to create a hierarchy perhaps, but in the end this is all calculation, too. Humans lose to AI bots at go because they miscalculate. Bots occasionally miscalculate (even I have beaten LeelaZero in an even game by creating a ladder) but humans miscalculate several times in each game. Humans likewise lose to humans because they miscalculate - it has nothing to do with who knows the most proverbs or josekis except insofar as heuristics help us overcome the deficiencies in our calculating abilities.

If I'm wrong I'd like to hear why in a non-technical way I can understand, but just staying on topic for a change would be nice, too.

What do you think about these instead :

https://deepart.io/latest/

Go has a simple objective measure of success, so you can start with a seed of human games (or even no seed, like the AlphaGo Zero bots) and through reinforcement, create something that is better than a human. With art, our judgment is the evaluation function.

It doesn't follow that an AI could never create art better than we can. After all, just as artists surprise us by creating new things that we wouldn't have anticipated, so could an AI. But it does mean that it's a harder problem.

_________________
Occupy Babel!

Putting Deep in the name doesn't make it AI in the neural network sense.



Well, I can accept AI could create something that passes the Turing test for art but it would have to more than something just new and/or surprising.

The main thing for me, though, is that you agree it's a harder problem than go. I suspect go players would love, and will strive, to find reasons to say go is harder, but I think we need to be realistic. I always remember Michael Redmond saying go is actually a very simple game. We just have to stop making mistakes.

John, I was trying to address the point you made in my post. Sorry it didn't come across clearly.

I agree that the painting is rubbish, but I don't think that's necessarily because computers can't learn to make aesthetically pleasing art. I think it's because the group of artists behind this project weren't trying to make art that you or I would like - they were trying to make something that would grab a lot of attention. The shortcomings of this piece of art make it more likely to do so, not less. My point about the money wasn't that art that sells for a lot of money is inherently good, it's that it may have met the goals of the person who made it. In that regard, I don't think that algorithm failed, and I don't think the result indicates that a computer is incapable of generating good (by classical standards) art.

I do think that making a neural network that can generate novel, quality art is likely to be a lot harder than making a computer program play go. A board game is a very well-defined problem space; art is not. And I doubt that the techniques used to create go playing software is the best way to go about making an art creation program. I don't even think we'd need to use anything resembling AI - a modern computer could use a 3D scanner to make high quality sculptures of a model or easily transform a photograph of a scene into something that resembles a watercolor painting. Using AlphaGo like neural networks to "learn" from human paintings is just a gimmick.

I think the big problem with applying "AI" across various domains is that it's largely a misnomer. We have developed an algorithm that is very, very good at learning to play go (and many other things). It's not really intelligent in the way that we normally think of the term, but when we call it "artificial intelligence" it's easy to ascribe properties to the software that just aren't there. I do think the capabilities of software that can learn by example from large data sets instead of being fed hard coded stimulus response behaviors is very exciting, but there's no reason to think that algorithm is the best approach for all fields nor reason to be dismissive of older approaches that can provide excellent results in their respective domains.

It's good you make these excellent points, jeromie, to remind us all to be both more precise and less gullible.

But from your more expert perspective, can you go so far as to say go AI is essentially a matter of better calculation? And if so, can we infer that AI bots will teach us little beyond needing to learn to calculate better? (I think that sums up the chess experience.)

The point of generative art is not to try to make fake human art, it's to make something else entirely (which happens to elicit responses that are guided by how we have learned to approach actual human art). If you think that that is bullshit, that is a valid opinion and you are totally entitled to have it, but it doesn't seem like there's much point in continuing to debate anything else about the subject, since we're starting from a different set of axioms.

That is true, but in this case the pictures linked to by Tryss were in fact generated by deep neural networks.

I like them.

The major difference between these and the top post exemple is that it's reproducing reality in an imposed style, instead of having a fuzzy goal like "draw me a portrait in the 14th-20th style"

To adapt a quote from the excellent movie The Princess Bride: All computing is calculation. Anyone who says differently is selling something.

However, whether (or what) we can learn from it is a question about humans as much as it is a question about computers. Humans are very, very good at recognizing patterns. We're also very good at coming up with our own stories for why those patterns are there - whether you look at mythology, science, art criticism, or any other field you'll find people putting their own interpretation on the patterns they find in their object of scrutiny. I think that humans are likely to pick up a lot of patterns from the new AI algorithms, and we'll make our own stories for why they work. (Sometimes our explanations might even be right!) I think the algorithms being used are more likely to generate patterns that are useful to humans than the older approach that was used for chess. We'll never catch up to the computers because they are so much better at calculation than we are, but that doesn't mean that all we can learn from them is the necessity of calculation.

The recent thread about the value of Shusaku's ear-reddening move is a good example. Many humans have learned from that game (and that move in particular) because we attached a story to why it worked. It turns out that it likely wasn't the best move after all, but many of us have still taken to heart lessons about looking at the whole board when choosing a move, nullifying your opponent's plans, etc. Do we really understand why the new joseki that have been adopted after AlphaGo work? Maybe, maybe not. But they are still useful patterns, and humans can become better go players by incorporating them.