Dream Denoising Model: Generative Artificial Intelligence as a theoretical model for understanding dream-building

Abstract

Artificial Intelligence (AI) has been rapidly developing in recent times, and with it, new ways of understanding psychological constructs based on their underlying mechanisms. The purpose of the present article is to propose a theoretical model for phenomenological dream-building based on non-equilibrium statistical physics, particularly through core processes observed in Latent Diffusion Models and Generative Artificial Intelligence. Drawing upon the Integrated World Modeling Theory (IWMT) of consciousness, diffusion models and current literature about the phenomenology of dreams, the present model (henceforth, Dream Denoising Model [DDM]) argues that dreams are the result of a denoising process by which the Dreaming Brain System (DBS) resists entropic states caused by impaired information integration, which reaches its lowest during dreamless sleep or Slow Wave Sleep (SWS). To do so, the DBS relies on memories (dataset), conditioners such as Stable Traits and Transient States of the dreamer (prompt) and predictive processing (cross-attention) in order to generate and predict an internally-generated model. The article also proposes the notion of a Denoising Circuit by which the repetition of sleep cycles between alpha waves and slow waves enhances the DBS ability to generate and predict more complex, vivid and bizarre oneiric experiences.