描述
I've long been frustrated by the literalness with which fields like computer science and computer vision approach the topic of seeing. It's a neat trick to point a camera at a picture and have the caption "hot dog" or "not hot dog" appear on the screen. But if you show Rene Magritte's iconic "The Treachery of Images" to such an object recognition model, the classifier will invariably return the result: "This is a pipe." Something is wrong here.
Visual perception is squishy and slippery, formed by each of our unique biological makeups, our memories, history, culture, and our own subjectivities. Just as a monarch butterfly sees a flower entirely differently than a field mouse, a medieval Spanish farmer sees a comet in an entirely different way than a contemporary architect. An early 20th-century psychoanalyst or semiotician might understand images from a dream quite differently than a present-day cognitive neuroscientist. "Seeing" is a deeply historical, cultural, subjective, and even political affair, profoundly shaped by our sensory and social environs.
With the "evolved hallucinations" project, I wanted to see what would happen if we tried to build computer vision models based on a wide range of historical, cultural, and notional worldviews. I began training models on allegorical art, symbolism, and metaphor, using image-vocabularies drawn from literature, philosophy, poetry, folklore, and spiritual traditions. Could I build models that embraced the slipperiness and squishiness of visual perception? What would it mean to build a model designed to "see" the world through the extended allegory of Dante? What might the world look like through the "eyes" of future seaweed on a post-human earth? Or the worldview of a Cassandra-like being, fated to see the future but helpless to change it? The Evolved Hallucinations is my partial answer to that question.
- Trevor Paglen
