TNG:  What does a typical day in the studio look like for you?

DS: I wish I had a typical day to be honest! I have a full-time day job as a UX designer, so most of my days are spent doing that. One of the reasons I gravitated toward machine learning was that the process is often hands off. I set up my training and every morning check on, start my work day and then check it again in the evening. Since I work remotely my studio is also part of my office, so sometimes I can sneak something in during a lunch break. The studio is filled with machines, some very high tech (machine learning GPUs) and some lower tech (pen plotters, basic motor systems). So a lot of my practice is making one thing in one way, then bringing it over another medium. Automation is large part of my practice, not necessarily because I’m lazy, but because it allows me to do things on a grander scale than I have the capacity for with my other daily requirements.

TNG: What is your creative process for starting a new body of work or exhibition?

DS: This is always the hardest question for me to answer. I don’t necessarily have a set process for starting new works. I have a backlog of ideas I keep in mind, and sometimes that lines up with the theme of a show or moment. But technology plays a huge part in my work, so sometimes I get access to a new technology or I learn a new technique that inspires a new project.

I make a lot of little demos and experiments that I share (mostly through Instagram). This is probably the most important part of my practice’s success because art directors and curators find me through that. I’m always surprised by what resonates with them because its rarely ever my finished work but usually one of these sketches. That sketchbook is something I often return to when starting new projects.

TNG: What artists or art works have inspired you the most in your career?

DS: Maybe I’ll talk about two projects that inspired this work, albeit somewhat indirectly.

Erik Nitsche’s Dynamic America. I have a large collection of designer Erik Nitsche’s work, but I think Dynamic America is easily accessible via used book stores and a real treat. It was originally planned as a film on the history of General Dynamics, but was eventually produced as a book. But it’s a book influenced by its filmic history—it flows the way a film might, includes sheet music, and features historic images moving across each page. It was a multimedia affair created in the late 1950s!

Jodie Mack’s Persian Pickles. A post-AI film if I ever had to define one, it’s one of many Jodie Mack’s works I’m inspired by. A cornucopia of paisley fabrics, its edited into a coherent animation that boggles my mind. A curated dataset and interpolation film that is done completely by hand!

TNG: Your artworks include a mix of materials used - elaborate on that?

DS: I guess I get bored really easily! I’m interested in the affordances and materiality of different media. While much of my work exists digitally, I think there’s something uniquely interesting in the interplay of physical and digital media. Robin Sloan has this really interesting theory about this interaction he calls “flip-flopping” that I’ve integrated into my practice. My work could not exist as purely analog or digital. The symbiotic relationship is the key for me. “Fossil Record: Etched in Stone” is a great example. Etching the machine learning interpolation into found 16mm film makes the digital analog, but it required me to still scan it back in and fix some of the issues with it to make the final piece. It actually looks quite terrible as a pure projection!

Mixed materials also allow me to explore sampling and how to create something new from existing material. The current debate about machine learning models “stealing from artists” is fascinating to me because I grew up on rap music and found footage experimental film. “Stealing” has always been a part of the art I enjoy, but its about transforming existing materials into a new expression. And for this particular series, I see the found materials add a dimension to the “archaeology” metaphor I’m exploring.

TNG: How did you come up with the idea of using a floral machine learning model for our upcoming exhibition?

DS: I’ve come to realize that a through line in my work is connecting machine learning to other scientific endeavors of the past. Many artists reference other artists’ works in their own, I think I’m referencing the history of science sometimes. The theme for the show, nature, made it obvious for me to use something from my natural image collection. But I also twisted the theme of this series as “human nature.” It’s about why we as humans are constantly trying to make sense of the bits and pieces we have available to us. In the early part of the 20th century scientists were finding dinosaur bones and trying to create some logical collection of them. But often they were very wrong. And that’s a large part of scientific history to me—the hubris of humanity and how often we’re very wrong despite our confidence. So the flowers, and their outlines, become a metaphor for that lack of data but our ability to sometimes make sense of it all.

I guess there’s also another reason to use this particular set of flowers. While the debate about AI and copyright rages on, this dataset is comprised of images from the Biodiversity Heritage Library who have done an amazing job of collecting rare biology books. The images are either fully in the public domain or ol enough I feel comfortable using them without infringing on a living artist’s body of work.

TNG: Could you explain the process of training your floral machine learning model? What kind of data did you use and how did you refine the model?

DS: I’ve been working with this particular dataset for a couple of years now, although this model was re-trained specifically for this series. Even the creation of the dataset took me a few months, as it involved taking images from the Biodiversity Library and cleaning them up and cropping them in a consistent way. Once I have the dataset—in this case over 6000 unique floral illustrations—I need to train the model. That takes one to two months using open source code I’ve slightly modified. Its not a super hands-on process, just time consuming. Its trained on a computer I built myself using four high end NVIDIA GPUs. Once I have the machine learning model its finally time to actually make the work!

Everything in the show uses an interpolation. Imagine wandering through a massive city and at every intersection you have a different flower. As you walk block to block the flowers morph from one to the other. That’s basically how an interpolation works. I set the intersections and the model does the job of creating these very smooth morphing animations. It can create these interpolations in a few minutes, so that’s the pay off of months of work. And there are quite literally endless interpolations in these models. It’s an unique aspect of these machine learning models, and why I use them over some of the more common text-to-image models popular right now.

TNG: What are the unique challenges and opportunities you encountered while working with machine learning in the context of creating art?

DS: The biggest consideration—especially in the process I use—is the amount of time it takes just to get the models that I’ll use. My current process uses StyleGAN-XL which takes about a month to train on my four GPUs. So I have to plan well in advance, and am often training models without any clear idea of how they’re going to be used beyond “I’ll figure it out when I get there.”

I teach a lot of other artists to use these tools and I find they often fall into two camps. The first camp, who want to have complete control of their output, tend to hate machine learning. The idea that you have to use some secondary source (text, for example) for image generation or that you have to generate hundreds of images before you get exactly what you want really frustrates this group. The other group, artists who maybe allow for chance and randomness in their work, gravitate toward ML because of its power and ability to create series or multiples of a concept.

TNG: How do you balance the creative aspect of art with the more technical aspects of machine learning? Is there a specific interplay between the two in your artistic process?

DS: I think all of my artwork balances some technical aspects with artistic intent. As mentioned previously, sometimes a new technical opportunity creates a project because of its particular affordance, but it's just as often I have a concept and am looking for the right technique to visualize it. And because I usually have crazy ideas that would be impossible for me to implement purely on my own, I often end up writing scripts or full on software programs to achieve what I need. 

Technology is cool because its reusable and scalable. Many of the techniques I use in The Fossil Record were made prior to this work, and often for very different reasons. The Fossil Record: Carbon Dating is a good example. Years ago I wrote a program to convert vector graphics in Adobe Illustrator to my very old pen plotter. It took me months, but I still use it to this day. Then last year I created a script to convert video frames into storyboards, so I can use that to create stop motion drawn animations by combining it with my pen plotter script. Five plus years of tooling made that video possible—I couldn’t have imagined that when I started.

But I think its also very important to separate technology from concept. My Instagram sketchbook is full of cool sketches, but they are often not concepts. I worry quite a lot about becoming a “tech demo” artist, where all I do is show off new code or techniques. But I want to ensure that my work also contains ideas important to us culturally. Right now a lot of my work is exploring the nature of modern technology because, well, I use that technology, I think about it a lot, and its a topic so many of us are grappling with right now. I don’t think I propose solutions or “right ways” of thinking about this tech, but rather try to share my own emotions about this world in this exact moment.

TNG: Do you see yourself continuing to explore the intersection of machine learning and art in future projects? If so, what areas or themes are you interested in exploring next?

DS: I’ve been pretty active with machine learning for 5 years now, so I expect that to continue. I’ve always seen ML as a tool in a larger art practice. I’m not really a purist that some AI artists are—I don’t care my output was made completely with AI. 

For me I’ve come to see the influence machine learning has had on my vision of the world. I’ve coined this term “post-AI” (it started out very tongue-in-cheek, but I think I’m using it more sincerely now) to describe how even the work I make that is not using machine learning is influenced by my machine learning practice. This often means I approach everything as a dataset, or break ideas down into components that become systems to explore. The Fossil Record obviously uses machine learning in it, but the overall concept is very post-AI. The output can be scratched into film, drawn with a pen plotter, or manipulated digitally using motion graphics techniques. But that underlying structure runs throughout.

TNG: Can you share any upcoming projects or exhibitions you are working on that our audience should look out for?

DS: A lot of my projects right now take the form of experimental films/animations, so I have work in a number of film festivals this summer. I just finished Floral Zombification via Attention Node Networks; it just premiered at the Winnipeg Underground Film Festival. I’m proud of that one; I think it's one of the earliest examples of a complete film using the new batch of text-to-video machine learning models. I just wrapped filming on a live action version of Selfie Song. Part of my post-AI vision is taking machine learning models back into the real world, and shooting this piece on 16mm film and editing it to feel like diffusion models is something I’m really excited about. I expect to have that film completed sometime this summer. And lastly I’ll have a new animation in this year’s edition of Kinomural based on my Editation on Violence series.

If you’re an artist interested in using machine learning and other algorithmic tools in your practice, I’ll have a couple classes available later this summer. I announce them on social media, my website, and my newsletter.