Human dreams represent the brain's highest-bandwidth generative process unconstrained by sensory input, social filtering, or logical consistency. During REM sleep, neural activity encodes creative recombination, emotional processing, and abstract problem-solving in a form that is beginning to be experimentally decodable. This paper proposes a formal information-theoretic framework for dream neural decoding: we derive channel capacity bounds for the fMRI-to-content reconstruction pipeline, establish theoretical limits on dream sequence fidelity, and specify the architecture of a generative model trained on reconstructed dream data which we call DreamML. We connect this framework to the emerging literature on neural decoding during sleep (Kamitani, Gallant, Dehaene), to the author's prior work on Shannon channel characterization of physical systems, and to the open question of whether subconscious generative processes contain learnable structure that waking-state data does not. We identify three testable predictions and four open questions at the intersection of neuroscience, information theory, and machine learning.
Matthew Busel (Thu,) studied this question.