In the summer of 1895, Freud had a dream about a former patient, now known as “Irma’s Injection,” that helped him understand the role of the unconscious in dreaming. According to Freud, our unconscious constructs dreams to express our deepest desires and then distorts those expressions to protect us. By this logic, a dream is therefore a form of wish fulfillment and suppression.

Although Freud’s theories have long been rejected, they provide a useful framework for thinking about why dreams and the unconscious are so often invoked to describe artificial neural networks and their psychedelic visualizations. Just as the latent content of a dream can never be fully understood, the architects of artificial neural networks do not fully understand the mechanisms of deep learning. Does this lack of understanding act as a form of dream distortion, protecting us from our deepest desires? When we build technologies that “dream,” what desires are we trying to fulfill, and why should we protect ourselves from fully expressing them? Artificial neural networks are modeled after our own and trained on the web of images we produce. Given that these models “see” with the same biases as mainstream culture, what do so-called “neural network dreams” tell us about our own unconscious dreams and desires?

Created using a deep learning text-to-image model, Fulfillment of a Wish experiments with storytelling in latent space, reflecting our own vibrant but nightmarish images back to us.