Data vs. Da Vinci

4.8.2019   |

10,000 tulips, the price of Bitcoin, and competing computers. On this episode of Fixed That For You we ask “Can a computer make art?” If that sounds too philosophical, just remember, artificial intelligence is trying to save us time by doing the repetitive jobs we hate. Artist Anna Ridler goes from writing simple formulas in Excel to directing a Generative Adversarial Network that creates a dynamic art installation.

Show Notes

See how Anna uses AI to bring “tulipmania” into the future with her latest installation called, Mosaic Virus. You can also read more about this unique technique on how Anna brings life drawing and machine learning together.

Curious to learn more about artificial intelligence and how it views the world? Watch this video on “What the world looks like to an algorithm.”

Want to geek out more on GANs? Here’s a beginner’s guide on everything you need to know about generative adversarial networks.

Read Karen’s latest published articles here.


Cara Santa Maria: This is what it's like to be an artist today.

Anna Ridler: I set them to train overnight, so I'll hit go and then wake up in the middle of the morning. It's like being a kid at Christmas, when you run and turn on the computer and you're like, "Is it going to work?"

Cara Santa Maria: Welcome to a world where artificial intelligence creates art.

Anna Ridler: This kind of algorithm is producing an image using pieces of all of the knowledge that it knows.

Cara Santa Maria: This is Fixed That For You, an original podcast from Segment about solving problems with data and algorithms. I'm Cara Santa Maria. In this episode, Anna Ridler takes a massive dataset and tries to turn that into a work of art.

Anna Ridler: I had no idea what it would produce.

Cara Santa Maria: It's the story of a young new artist establishing her voice in the face of long-held traditions.

Anna Ridler: I think this idea of drawing and painting being a human thing will lessen.

Cara Santa Maria: In 2003, Anna Ridler was a teenager living in England.

Anna Ridler: I was really interested in painting and kind of representation in a very kind of realistic way. It was quite traditional, how paint works and how drawing works.

Cara Santa Maria: She went to art school to pursue her dream but it wasn't quite what she expected.

Anna Ridler: Because I was so young and I think you're putting yourself out in the world and you're saying, "This is what I believe." I think good art is kind of a little bit contentious. When you're 18 or 19 or 20, I think the type of person I was, I was very shy. I didn't really know what my position on a lot of things was so I found this thought of going to art school and having to do all of this quite scary.

Cara Santa Maria: What Anna is talking about is finding her voice. It's impossible for an artist to share their message with the world if they don't know what that message is yet. Intimidated, she left art school, got a degree in English Lit, and this led to a series of administrative office jobs, but she still found ways to be creative.

Anna Ridler: One of the things that I got quite interested in was actually through using Excel and being able to kind of mechanize things and to be able to mechanize processes. You don't have to do these very long, boring things. You can kind of create a little mini algorithm.

Cara Santa Maria: After learning Excel, Anna took some basic coding courses.

Anna Ridler: To begin with, I was just doing really simple JavaScript stuff. As soon as I started using Python, then I could start kind of making things.

Cara Santa Maria: Coding became her hobby and then her passion.

Anna Ridler: The idea of data and what data is didn't really exist for me when I was both an art student and at university studying literature. It was just kind of the idea that you can break the world up into tiny pieces, and then use those tiny pieces to then maybe reveal bigger, wider themes.

Cara Santa Maria: And it started to change the way she saw the world.

Anna Ridler: When you code for sustained periods of time, you become, I suppose, hypersensitive to precision of things. In a way, it is like anti-poetry because in poetry you're trying to build in as many possible meanings to one word, whereas in code you're trying to strip out any ambiguity. It's kind of like I have these two battling sides of my personality.

Auction voice: $190 ... There it is, $200,000. … are you in or out? I have $200,000, front left there. It's your bid, sir. At $200,000...

Cara Santa Maria: Around the same time Anna was undergoing this personal technological transformation, something similar was happening to the art world, the introduction of a subset of artificial intelligence, something called a GAN.

Karen Hao: GAN stands for a generative adversarial network.

Cara Santa Maria: That's Karen Hao. She writes about artificial intelligence for the MIT Technology Review.

Karen Hao: It is referring to two neural networks that are actually competing with each other.

Cara Santa Maria: Neural networks are computer systems modeled after the human brain. They can learn.

Karen Hao: They are able to do some pretty sophisticated math to transform data from one thing to another.

Cara Santa Maria: If you pit two neural networks against each other, this is what happens.

Karen Hao: One neural network is called the generator and it is in charge of creating whatever you're asking it to create. Let's say you're trying to generate dog images. The generator starts by training on all these dog images and then creating a dog.

Cara Santa Maria: Based on the dataset you give it, the generator will compose its own image of a dog.

Karen Hao: Then, the other network is called a discriminator. The discriminator will then look at the dogs that the generator created and, based on the same set of training data, it will determine whether or not it's a dog. Any time the discriminator says, "This is not a dog," the generator has to go back to the drawing board and try again. These two networks will compete back and forth, back and forth, until neither of them can really outwit one another anymore. At that point, you get a pretty darn good picture of a dog. That's why, if you are an artist and you're working with GANs, you can start creating pretty hyper-realistic images of whatever you want, whether that's a photorealistic picture of a dog or an artistic rendering of a Van Gogh image.

Cara Santa Maria: That's an important point. The GAN will create images of real dogs or of Van Gogh-style paintings depending on what you put in your dataset. When it comes to AI art, the composition of the dataset is everything. That is especially true now that it's starting to sell at auction.

Auction voice: ... Ahead of you. Selling for $480,000 then. Sold to panel 90.

Cara Santa Maria: In October 2018, an original painting created by a GAN was sold at Christie's for more than $400,000. It was the work of three French artists that operate under the name Obvious. They selected 15,000 portraits from the 14th and 20th centuries, fed them into an algorithm, and voila. This amplified an ongoing debate in the art world about the role of technology. Can machines do something inherently viewed as a human skill and, even if they can, should we call it art?

Anna Ridler: Yes, it is a human quality but ever since technology has existed people have been trying to use this technology to push what drawing and painting means. Even when you're looking at the Renaissance, people were using technologies like lenses and things like that to push how they draw and how they painted. You're just starting to get to the point where machines are able to produce drawings that could be seen, and in fact have been seen, as being better than those produced by humans. I think what's interesting is how you can kind of not see that as a threat and how you can kind of embrace it and then use it as a way of pushing practices.

Cara Santa Maria: After leaving art school without the confidence to express a vision, Anna decided to return to art with conviction on the less popular side of a growing debate.

Anna Ridler: Rather than just kind of thinking about art in a very kind of paint, pencil, canvas, paper kind of way.

Cara Santa Maria: She wanted to experiment with GANs and AI-generated art but, in doing so, Anna also wanted to distance herself from some of the other artists using machines, especially those who were using existing image banks.

Anna Ridler: That's fine, but it depends what you're trying to do with your work and it depends what things you're interested in exploring. I'm not interested in kind of taking the world as it exists.

Cara Santa Maria: Anna didn't want to explore someone else's dataset but the ones she would come to need would be huge. Anna found GAN algorithms on GitHub and modified them to her specific needs. She experimented with a couple of small projects. Then she was ready for a more complex one, something that would firmly establish her artistic voice.

Anna Ridler: I decided that I would use a GAN to create a tulip that would then respond to the price of Bitcoin.

News reading: Taking your eyes off the screen, don't you, for the weekend and you get another 13% bump higher in the price of Bitcoin ...

News reading: We've been talking about the rise of Bitcoin but Thursday's action was out of this world ...

Cara Santa Maria: Late in 2017, Bitcoin went crazy. It bounced up and down from over $22,000 to just $6,000. Anna saw similarities between the cryptocurrency bubble and a tulip craze that overwhelmed Holland in the 17th century.

Anna Ridler: A tulip bulb about that time, at the height of the bubble, went for the same price as an Amsterdam townhouse. I was really interested in how you can combine these two kind of moments in history and I wanted to do that using AI and machine learning, which is going through its own bubble at the moment.

Cara Santa Maria: It was a way to explore technology while honoring her artistic roots.

Anna Ridler: I wanted to create something that referenced old Dutch still lifes, so the still lifes which have the beautiful flowers and the black background, so that, when you would look at it, it would look like a tulip but it would be constructed by an AI.

Cara Santa Maria: For that, Anna needed a dataset of tulips, thousands of tulips. The easiest way would be to do the same thing the French artists Obvious did, an online image search.

Anna Ridler: You have access to every image possible but only if it's been tagged in the right way, or only if it is kind of coming up with the keyword search that you say. There are so many problems.

Cara Santa Maria: Not the least of which is bias.

Karen Hao: Some of the biggest, most popular open source image sets were compiled at a time when the majority of images on the internet came from the US or from Western countries.

Cara Santa Maria: That's Karen Hao from MIT Technology Review again.

Karen Hao: A lot of image recognition systems fail to recognize images that are not the standard Western image. An example, if you show an image recognition system trained on this dataset, a picture of a Western bride wearing a classic Western white dress, then it'll say "bride," but if you show it an Indian bride wearing a classic traditional sari, it'll just say "woman." Because a lot of people are using the same exact image datasets, whatever bias that dataset has has been trickling into all image generation, image recognition systems. If you want to do something very specific with your artwork and you want to express something very specific, then the best way is to just do it yourself.

Anna Ridler: When I make a dataset, I want it to reflect my worldview. I am building the world from the ground up so that, if it's making mistakes or repeating biases or showing or creating a world, it's my biases and my world. I can live with that. Cara Santa Maria: 10:39 For the tulip project, Anna's concerns weren't so much about bias as quantity.

Anna Ridler: If you Google "tulip black background," you will not get 10,000 different tulips on a black background. That just doesn't exist so that was something that I had to make and I had to do.

Cara Santa Maria: She set out to build the world's largest tulip database.

Anna Ridler: If you decide to take 10,000 photos of tulips, you need lots and lots of tulips. I went to the Netherlands and I would go to the flower markets and buy hundreds and hundreds of tulips.

Cara Santa Maria: Which earned her quite the reputation around Amsterdam.

Anna Ridler: The first time I went, I think they thought I was getting married. Then, I came every single Saturday for the next month. I think they just thought I was the crazy English lady.

Cara Santa Maria: There were no shortcuts for this. The dataset had to be built one flower at a time.

Anna Ridler: You have to take the tulip, strip all the leaves off, hold it, photograph it.

Cara Santa Maria: But the dataset she built wasn't random. It reflected her own artistic vision.

Anna Ridler: You kind of notice you need to buy more pink ones or you need to buy ones that look a certain way because you're getting too much of one color in your dataset just by going to the same markets all the time, so I'd have to go to different markets to try and get different types of tulips.

Cara Santa Maria: At the time of the craze, striped tulips were the most valuable so Anna's idea was that, as the price of Bitcoin rose, the tulips in her art would get more stripes.

Anna Ridler: I needed to have enough stripey tulips that it would be able to learn what a stripe was, and stripey tulips are actually really difficult to find because they're valuable. They're rare, they're not as common as normal tulips, so I was going around, running around the Netherlands, just buying as many stripey tulips as I could find.

Cara Santa Maria: Week after week, she shopped, pruned, and photographed, but, even after taking 10,000 tulip photos, she still didn't have a dataset.

Anna Ridler: I have all of these photographs but, to an algorithm, it doesn't know what a tulip is, it doesn't know what color it is, it doesn't know what a stripe is, it doesn't know whether it's a bud or dead. To it, it's just kind of like ... It's just maths. I had to go and then hand label each of the photographs so that it would be able to understand what was in that photograph.

Cara Santa Maria: This was the second time Anna's own bias impacted the data.

Anna Ridler: You'll have something that is pale, very pale pink, and you're like, "Is it pale pink or is it white? Is it yellow or is it orange? Is it pink or is it red?" As soon as you're confronted by putting something into a binary bucket, it suddenly becomes very difficult. This is just for something as easy as the color of a flower.

Cara Santa Maria: It took Anna six exhausting weeks to label the flowers. Then, she finally had something to feed into the adversarial network. Then, more problems.

Anna Ridler: I finally have my dataset and then I start to run it. The first time I run it, it doesn't work. It just doesn't. It just produced just green objects. Oh, I just cried. It was just horrible because also I'd just spent months making this dataset, and I knew I was having to exhibit it so it had to work.

Cara Santa Maria: She rewrote a few sections of the code and tweaked the color variables. Then, she gave the data back to both the generator and the discriminator to start creating tulips.

Anna Ridler: It was really, really hard to condition it, to get it to understand what a stripe was, because some of the time it would just not produce anything. You were having to test whether it could understand this idea of conditioning by doing it by color, and then kind of then adding an extra layer of kind of stripiness to see if it would understand. Then, sometimes it would get really blurry, and sometimes that would mean I mislabeled my data.

Cara Santa Maria: It's kind of like teaching a child to draw. You give them an example to follow but encourage them to be creative. When they stray too far from the objective, you alter the instructions to bring them back in line.

Anna Ridler: I ended up making probably about 30 to 40 different models to try and make it work.

Cara Santa Maria: This isn't something you can run on a laptop. The processing power required is massive and expensive so Anna had to use cloud-based computing. She was making progress, but then she came up against another major obstacle.

Anna Ridler: There's something called mode collapse which happens, which is when it over-learns or it kind of under-learns and it just produces mush.

Cara Santa Maria: Mode collapse. It's a recurring problem for anyone training a neural network. It's a case of too much feedback from the discriminator. Going back to that analogy of the child, it's like they get really good at stick figures and they get so much positive feedback they just keep drawing the exact same picture over and over again. It can also go the other direction, when the young artist gets too much criticism and they give up trying and just start scribbling. But Anna considered mode collapse part of her art.

Anna Ridler: It's something called the learning rate, so you can kind of see how well something is training, how well it's understanding. It gets better and better and better and then it will suddenly collapse. That echoes the subject matter. That's kind of how the stock market works during these kind of speculative bubbles. It goes up and up and up and then will crash.

Cara Santa Maria: Unlike the stock market, Anna finally got past mode collapse by tweaking the algorithm and the dataset again. On her final attempt, she set the GAN to train overnight.

Anna Ridler: Is it going to work? Has it worked this time?

Cara Santa Maria: She woke up the next morning not sure what would happen.

Anna Ridler: It starts off with nothingness and you kind of hit the space bar to see it progressing, to see it learning. You kind of see it kind of going ... You see these tulips forming from being really, really blocky. Then, you keep on going and it's training more, and then they start to become real looking tulips with stripes. Then, you're like, "Oh, my God. This might actually work." Then, it's ... You see it and they become more and more defined and more and more tulip looking. Then, you realize that it has actually worked. It's amazing. Yeah, it was just ... I was just so happy.

Cara Santa Maria: The algorithm was actually drawing unique pictures of tulips on its own.

Anna Ridler: Because I have such an intimate knowledge of the dataset, I can tell that it's made tulips that don't exist in the dataset but could be real tulips, which is really, really lovely.

Cara Santa Maria: At the same time, Anna thinks the GAN was following the same process as the Dutch masters when they painted still life.

Anna Ridler: They wouldn't just kind of have a bunch of flowers in front of them and paint from it because, if you look at those paintings, there are flowers from winter and spring and summer and autumn all in the same bouquet of flowers. Rather than kind of painting flowers in front of them, they would kind of use all of the flowers that they know in their mind and paint. They're painting from their imagination and memory rather than painting from what's in front of them, which is exactly the same way that this AI is working.

Cara Santa Maria: But there's one thing the Dutch masters probably wouldn't have added to their paintings.

Anna Ridler: Then, you start to make it dream.

Cara Santa Maria: Yeah. They didn't have Bitcoin back then.

Anna Ridler: There's a separate kind of script that you run which adds in the price of Bitcoin. Then, it starts to morph and change according to that.

Cara Santa Maria: The work is called "Mosaic Virus." What you see is a grid of 48 different tulips against a black background. They're constantly morphing from buds to full bloom and shifting colors from plain to striped. Each flower develops in its own way but, as a whole, the collection flourishes as the price of Bitcoin goes up and diminishes as the price falls.

Cara Santa Maria: As impressive as the final product was, Anna was still thinking about the dataset used to create it.

Anna Ridler: For me, this is such a huge important part of the project that now I display the database as its own separate work. The reason why I chose to do that is because I think people don't understand what is in a dataset. This is a huge amount of information.

Cara Santa Maria: "Mosaic Virus" and its dataset were unveiled to great interest at the Impact Festival in Holland. As with other pieces, it contributed to the debate over AI's place in the art world. You and I might have different opinions about whether a cubist Picasso is innovative or just a jumbled up face, but we all agree on where the credit lies. When art is produced by a GAN, it's legitimate to ask was it the artist or the algorithm?

Anna Ridler: There's a sliding scale. You have people who are just clicking and downloading a dataset that already exists and clicking and downloading an algorithm that already exists and putting them together and then just pressing a button. On the other hand, you have people like me and other artists who are really considering about what they're trying to do and how this is then reflected in the whole process that they’re making. There is a huge amount of labor and there is a huge amount of time that goes into it. I think that is what makes it art and not just a cool tech demo.

Anna Ridler: For me, I found it a really kind of liberating way of kind of exploring kind of how I process the world and what I make of the world.

Cara Santa Maria: If you want to see Anna Ridler's "Mosaic Virus," check out the show notes.

Cara Santa Maria: That's it for this episode of Fixed That For You, a podcast by Segment about solving problems with data and algorithms. You can find us at segment.com/podcast, plus subscribe at Apple Podcasts, Google Podcasts, Spotify, or wherever you do that sort of thing. We drop a new episode every two weeks.

Cara Santa Maria: I'm Cara Santa Maria. Thanks for listening.

Episode 7 - Data vs. Da Vinci

Episode 7 - Data vs. Da Vinci