In short, I create nonsensical titles for artworks by feeding a large dataset of text into a Recurrent Neural Network, and then use photography and Midjourney to visually interpret the generated text.
The dataset I use comes from the BBC’s online library of 33,000 sound effects, which can be viewed here: https://sound-effects.bbcrewind.co.uk/search
By downloading the CSV file I was able to isolate the descriptive titles of all of the sound effects, and feed them into the RNN. A few examples of titles I feed in - ‘Miners chatting in pit baths’, ‘Indisposed chicken’, and ‘Cod machinery’.
By downloading the CSV file I was able to isolate the descriptive titles of all of the sound effects, and feed them into the RNN. A few examples of titles I feed in - ‘Miners chatting in pit baths’, ‘Indisposed chicken’, and ‘Cod machinery’.
The RNN learns patterns within the text and then generates text one letter at a time. It produces contextually relevant text but also creates grammatically incorrect and nonsensical sequences due it’s focus on character-level patterns, lacking high-level language comprehension.
This is why it’s easier to get more suprising results using a narrow-focus RNN rather than an ‘intelligent’ Large Language Model like ChatGPT.
When setting the parameters within the RNN, ‘low temperature’ searches require less energy and involves reaching for the most probable answer. ‘High-temperature’ searches require more energy because it involves reaching for less likely answers.
These high-temperature searches might be inefficient and incur a high rate of error, but they occasionally return answers of surprising beauty and originality. They are more magical. A few examples:
Hoods Pausing, Tender Roles
Bruised and Subdued (3C3)
Wirht No Hodstandt
Restaurant Pre-Pause
Knight at Night-Bag
Dog Thrown Plaster
New Doo Sputter, 12”
Punter Turning
Small Floor
Coinstench Eton 196h
Pulses: Yord's Subdubtic
1 Fly Atmosphere In-built
Hothounds Formula Bangle
Subdued Conversation Salad
Comedy Gold, Before Scream
Pet Collapse, Fism Starts (Close Perspective)
Then I interpret these titles using a big blend of my own photography, found imagery and AI generated images.
Here’s a very basic diagram laying out the procedure, with my role as an artist highlighted underneath.
