Autobiographical paintings

Training an artificial intelligence algorithm on an autobiography

Summary
I wrote a 700-page autobiography and used it to train a machine learning algorithm. The algorithm generated a new text – an alternative, unstructured version of my life. I cut up sentences from this new text, and transposed them into paintings. I wanted to know if the algorithm's randomness could create poetic arrangements from my own life. On a conceptual level, the process aims to echo our tendency to narrate and curate our lives – sometimes to please algorithms.

Self-portrait (two truths and a lie)

Quantum nude in the short-term rental

The project manager looking at the sea

Cats and dogs in the automated warehouse

My mother playing a free-to-play puzzle game

Background
The starting point was to explore controlled randomness by merging old techniques (cut-up, hand-painting), modern tools (artificial intelligence, design software) and a very personal dataset (an autobiography).

Study 1 and 2,  30 x 40 cm, acrylic on cotton. Those are preliminary studies made with placeholder text, before I started writing the autobiography

Writing
I wrote my autobiography for months, chronologically and in English. I focused on mundane descriptions (events, people, places, objects, feelings), trying to stay away from poetry. Training a model requires a large dataset, so I wrote 700 pages (2MB, 365672 words). I used Max Woolf's textgenrnn.

As a performance, I filmed myself during the 200+ hours it took to write the first version of the autobiography. I made a 5 minute edit , in which I took half-a- second of video every time I turned the camera on.

Example of input text

"The best memories I had of this TV was watching Tintin on it. Every Tuesday night, a cartoon adaptation of Tintin would air. The style of the cartoon was identical to the comics. I was a big fan of the comics. In the cartoon, the stories were the same as in the comics. I was excited and avidly looked forward to Tuesday nights. My sisters were less interested in Tintin than me. My sisters were also younger. As a result, I was the only child allowed to stay up and watch Tintin."

Example of output text

"One of them were extremely there. She was clear that my age was full of smart and straight to your mother. You were in living in the morning. This antivation is in my mind. You are a poster with projects. I was much less into the house. I was happy and impressed. It was a child, but I didn’t know eating obscure tastes. Some other family entertained people. The musician hornet loved me. I was not confident, by ruddy as a sledge. He made a good job, and slept in the administration."

Painting
I read 700 pages of generated text. I cut-up the parts with words juxtapositions that seemed to produce new meanings (e.g. a sunny consultant). You can read 100 pages of the generated text here. I painted them on a canvas at random – without pre-determined layout.

Untitled, 120 x 80 cm, acrylic on linen

This painting felt too unstructured, so I looked at classic subject matters: still lives, nudes, etc. I defined 10 subjects matters – a small number, to contrast with the 700 pages of the autobiography. Most of those subject matters refer to sensations, like being caught in an anxiety-inducing email chain, or wondering about the parallel lives of a short-term rental.

I sketched preliminary layouts. They didn't feel coherent with the process because they lacked the machine's input. As a consequence, I messed them up with Illustrator tools that allowed for relatively controlled randomness (twirl, crystallize, etc). I selected the ones who embodied the most each subject-matter.

I painted these computer-generated shapes on linen canvasses (60 x 50 cm). I laid out text on them, and rearranged it with controlled randomness, diverting Figma's Autolayout. I selected the layouts that fit the subject matters. I cut-out the final text, and typeset it with IBM Plex, a typeface designed to "embody the unique relationship between man and machine". I hand-painted the text.

Conclusion
I tried to use algorithms like creative partners: I curated data for them (text, layouts), controlled their randomness to generate creative proposals (new text, new layouts), which in turn, I curated. This felt like working with laws of physics, that respond semi-predictably and gracefully to our inputs (cf. Walter De MariaChris Burden). In the future, I would love to see if algorithms can be trained to generate paintings in direct relation with our individual autobiographies. I wonder in what ways this would touch our hearts.

Credits: textgenrnn (Max Woolf) and char-rnn (Andrej Karpathy)