A game of construction - Helena Sarin

Transcript generated with OpenAI Whisper large-v2.

My name is Helena Seren. I live in New Jersey. I'm a ceramic artist and I live with my rescue dog, Cumin, my three GPUs, and my husband. And today I'll be talking about game of construction. Construction is a metaphor for art making by Pollock and he famously said, it is all a game of construction. Some with a brush, some with a shovel, some chose a pen, and some of us choose neural networks to do our art. But the first question, what makes me a ceramic artist now qualified for talking at a male conference? I will say that recently, not recently, a couple of years ago, I discovered this magical word, Kugei. And it means craft in Japanese, but literally it means engineered art. And since then I'm calling myself engineering artist. And as a Kugei practitioner, it was my whole life when I was coding the low level protocols in telecom, when I kind of expanded into machine learning seven years ago, and when I'm now a practicing artist making art for a living for the last almost four years. So Kugei is our word here. So, and I'll be talking about creating art, making art in the intersection of art and technology. So technology is our tool for us engineering artists. And it's not only with AI, it's basically for the last maybe century, it's even accelerated where technical invention brought the new art form. So we can pinpoint this in like history of art is like prints, then photography, cinema, computer arts, and now AI kind of like burst into the scene big time. So I work with generative adversarial networks. I mean, up until prompt-based AI came into the picture, I would say like AI interchangeably with GANs. Now I need to qualify that I work with GANs mostly. So the base recipe, I work, first of all, I work with my own data sets. I collect them. I make photography or I use my digitized sketches. So here is like my famous vegan model, which I took like salad leaves at the beginning of COVID basically sitting at home in lockdown. So I took, like I made two salad leaves, 2000 pictures of lettuce in like on my force for two days, making pictures in the light box of few boxes of whole food, organic, clean lettuce leaves. And I trained the network StyleGAN and it started after like few days of training, it started generating this, creating this generative produce, what I started calling after that, like Entusys. So this talk, I'm quoting a lot of, is a culture trash and I'm like always into quoting. And I think this quote by Carpate is really kind of like relevant for artist engineers who like team caring. So basically training your network is you sitting through the like days and nights watching your model kind of making stuff, saving checkpoints. And the difference may be with like researchers or scientists who work with neural networks. My loss functions are more kind of like wandering functions. I mean, I don't kind of care for particular loss. I care more about visuals and that's why I'm glued to the screen because I need to see like particular checkpoints where I kind of discover some gem. And I used to call training networks is my meditative regime, breathing in and out with the loss function. So but basically it's like a game of curation. I mean, a lot of this and for practitioners being them engineers or artists is just like you curating your results. And that's why I have this famous Burrows quote because you need, I mean, in any art form, you kind of like do sketches, you do a lot of curation. It's not kind of like particular to making art with the eye. But also I think it's like a great tool of honing your visual perception. I mean, I kind of like some of my early work is a cringe, but some of them is sort of like you know, looking at the world like a child and spoiled and in awe of what you see. And this is one of my basically one of my first publicly shared pieces, which I'm still very proud about. So game of subversion. I keep saying that, like, in a sense, us working with the same networks that researchers work, it's a little bit kind of like subverting what the model do. Because if researchers, their goal at large, getting to some digital correctness, in my case, is more like seeing interesting and novel and exciting stuff. So that's why maybe nobody cares about like first cycles of training. But for me, and this is a picture on the left, the fleeting beauty of early iterations. But also like when you get to like beyond the expiration date, I mean, I have some models on the right you see the example of model drop, of mode drop, where I trained basically the network for almost like two years on and off. Yeah, and this is the network that was trained on my sketches. But also one of the technique I used a lot after I kind of like passed this excitement of like something in something out, I started building the pipelines. So kind of like one of the example of like poor woman super resolution where I sort of like trained the next network or the result of the previous one, just getting to the resolution which I wanted or needed because like the initial result was like 256 by 256. So I love this quote by Carpati, who said, it is not a single model, it's a beautiful composition of many sub models. That's how my pipelines look because I have like about 100 models that I use practically all the time. Another thing that happens is like you make art in on the limitations of your medium. So early on, like again, to overcome this problem of low resolution, I started working with the grid. And for me, it became like a neural aesthetics in a sense. So I could have like interesting images and augment them or kind of like make them bigger by employing the grid approach. And then I was surprised where like famous artist Jack Whitten called this the grid is a DNA of visual perception. It's an amazing quote. I didn't think in this terms, but that's what it is. Another thing like again, like playing this subversion game, people who worked with neural networks early on remember this like problem of checkerboard. So kind of I jokingly started calling this checkerboard amplification series. So it's kind of like post-genism. I started writing a lot of like Python scripts, just doing like this computational grids playing with a few kind of like digital collaging and stuff. So that's how this series of non-pintilism and checkerboard amplification series were born. Another thing like for a couple of years, I worked with my sketches and it's resulted in this, what I call carvings of Latinsco, riffing on the Lascaux caves where like caveman's like our predecessors in a sense drew on the cave. So I think like we early AI practitioners drew our part on this Latinsco caves. But also I like this quote by Linda Berry who said like take advantage of basic human inclination to find meaning in random information. So basically the semi abstract output of the early models invited this kind of like planning and gameplay. And I worked with it a lot. I mean, I kind of like, it's sort of like honing your verbal perception or your kind of like pun mechanism. This is kind of like controversial, but for me it's very important. I kind of like burnt a lot on this. So when like almost everybody, I started with AWS of course, but then when I like five, six years ago started training on my own art, I decided I need my own setup. Not only because I didn't want to move like my data all the time back and forth and forgetting to disconnect the volumes. I decided I basically want my own rig. And I started with the GTX, GeForce like 1060, I think. And then kind of people of Nvidia gifted me with this quadro. And like I say, GPU is a girl best friend, of course. So now I have like this small stable of GPUs. And like, since I sort of like make this phrase in the different applied arts, I last year I kind of got my own kiln and I have like other kinds of stuff. And it's not about bragging. I mean, it's like printing. I used to print a lot and like outsourcing, it's a pain and you pay dear because first of all, in quality, because you can't control this. And another kind of like you pay money for food to people which you can kind of like not pay and make your art more affordable in a sense. So this was my star slide in every talks for the last like three years. And it sort of became obsolete with like prompt models and I'll explain why. So in a nutshell, what I was talking about here and it was relevant when like for training on your own data. So the more you own and the more you curate your data, the more original you are. So like in the origin of this chart, you have the less original art. So it's like art breather and first kind of like public trained applications. But then like prompt text to image applications came into for this year and we all basically gathered around this push the button thing. And this is like I thought I looked at this slide and thought like maybe it's not relevant anymore. And then I think this Z-axis with ideas and I think like since when you can find them, they join them. So we are all in the realms of this prompt models. I mean, it's inevitable and this is something the reality that we can't deny. So I think it's still kind of like relevant graph, but with more access, how ideas become more and more important. So that's probably again, I mean, something that take away from here. So to finish this part is just how I see this AI artist or basically any artist who works with like technology. So you're basically curiosity driven. I mean, it's like the world of what ifs you try this. I mean, what if I kind of like make this data set or this data set? You improvise on the spot. You see your model behave this way. You kind of like encourage it to do it or kind of like do something else. You are very attentive and good listener because I think like patience is a very important trait for working with technology. You need to be attuned to what you do, but you also need to like tinkering. It's a lot of tinkering when you write your code, whatever like AI or any other thing. So creative coder goes without saying. You make your own tools. I mean, the less applications you use, the more interesting you are. It's not guaranteed, but still kind of like a precursor. So collector process oriented, we already talked about this meditative qualities of AI training. Collector, I mean, I basically save everything. I mean, you never know. I mean, whatever, checkpoints, like outputs, you never know what you can use. I used to kind of like pre-curate them. Then I stopped doing it because something that looked like a junk two years ago, all of a sudden becomes trash right now. So you never know your perception changes, the tools you're working with changes. So a lot of like dynamics. Also easily excited and easily bored. I mean, because we work with novelty tools, you move around a lot. So you get excited, you try it. I mean, you maybe need to stay with certain kind of like stuff for a while to see. So you're not jumping from thing to thing like without giving it a chance to express itself. And also I think it's very important not to be precious about your work. And of course, having a good sense of humor helps. I mean, because it kind of like helps with the frustration, but also kind of like because AI is still kind of goofy often. So in the next like five minutes that left, I'll talk about projects that I worked at. And I mean, neural bricolage manifest, like I keep asking, getting this question, what bricolage means. And neural is kind of like comes from neural networks, but bricolage is basically making useful and beautiful out of heap of latent space. I mean, this is how I kind of like riff on the definition. And my art is for people. So again, like artists, some artists say, oh, I will create because it's like I create for myself. There is a part of this, but they make stuff for people. Without people, I would be just curl in the armchair and just reading stuff. So a little bit of like, this is like mandatory milestone line. And how I call myself like taking this expression from my friend. So I basically awkwardly stuck between digital and analog. In fact, I was very surprised when somebody like three years ago called me a digital artist. I never consider myself being it. Digital was always means to an end and then being like projects I discuss right now. So I mean, I try to kind of like do interesting stuff, different stuff with the output. So I went through the period of printmaking, silkscreen, cyanotype. Then I started making the artist books because it's so kind of like I love books and I think it's like everybody needs to make at least like one book in their lives. So the first book was like a religious trope on like this Genesis play. And it was like small, all my books are self-published. So this was like my first books with the latent doodles. Then I think it's very important to have like collaborations with like artists. I'm very independent, but still I believe like when you find the great collaborator is just like amazing. So in this case, Dmitry Chernyak, a generative artist created for me the training data sets with 3GS. And then I trained like 17 models total and so like from scratch. And that's like on the right, the results that we are publishing the book is coming up soon. So another kind of collaboration is the book of vegan was just published with the amazing food writer, Kate Ray, who wrote recipes for this book. I'm very proud. I designed this book from scratch. It's like I went overboard. It's 150 pages thick, but still this is what it is. And like we discussed, I'm in the pottery for this year. And my first foray with generative pottery was like a couple of years ago. And I liked this like riff on pottery baron, which became pottery again. And yeah, I mean, then I kind of like stumble on a learning curve because it's a slow process. I'm kind of like all over the place person. So ceramics was not that, I mean, it was tough. But again, COVID and stuff sort of slowed me down. And I think like this time I'm ready. And basically I have my first fruits of my kiln coming out. And again, in a nutshell, it's something that I use again outputs to kind of like go through the convoluted process of printing 3D, you know, molds, then pressing or slip casting. So it's like very involved process. And maybe it's such a difference from like fast paced TIR that I really kind of like getting very used to it. And I love this. But to finish it is all a game of construction. Thank you.