Machine Dreaming and Perceptual Hallucination

Sep 27

Diffusion models turn noise into images. Dreaming brains turn noise into narrative. The comparison is closer than it sounds, and more instructive about perception than about machines.

This post is an essay. It draws on established findings in neuroscience and machine learning, but the parallels it proposes are interpretive, not demonstrated equivalences.

Machine learning borrowed the word "hallucination" for model errors, and critics have pushed back ever since. Confabulation is more accurate for language models: fluent, confident output unanchored to fact, produced without any intent to deceive. But the original metaphor points at something worth taking seriously. Generative systems produce structured content from noise and learned priors. So, during sleep, do brains.

What dreams might be for

There is no settled theory of dreaming, but several accounts have serious support.

Allan Hobson and Robert McCarley's activation-synthesis hypothesis, from 1977, proposed that REM dreams arise when the forebrain tries to make sense of largely random brainstem activity. The narrative is imposed after the fact on noisy input. Later work complicated this, but the core image — a pattern-making system applied to signal with little external constraint — has proved durable.

Memory consolidation accounts point to replay. In 1994, Matthew Wilson and Bruce McNaughton showed that hippocampal cells in rats that fired in sequence during waking exploration fired in similar sequences during subsequent sleep. Sleep appears to reactivate and reorganize experience, integrating it with existing knowledge.

Erik Hoel's overfitted brain hypothesis, published in 2021, drew the machine learning connection directly. Hoel argued that dreams function like the noise-injection and data augmentation techniques used to keep neural networks from overfitting. Daily learning is narrow and repetitive; dreams inject strange, corrupted variations of that experience to keep the brain's model general. The hypothesis is testable and remains under investigation.

The machine parallels

Each account has a counterpart in generative systems.

Diffusion models, developed in work by Jascha Sohl-Dickstein and colleagues in 2015 and made practical by Jonathan Ho and colleagues in 2020, are trained by gradually corrupting data with Gaussian noise and learning to reverse the process. At generation, they begin from pure noise and iteratively denoise it into a coherent image, guided by learned priors about what images look like. Structure is not retrieved; it is imposed on randomness by a model of the world. That is activation-synthesis in engineering form.

Replay has a direct analogue in continual learning. Generative replay methods, proposed by Hanul Shin and colleagues in 2017, train a generator to produce samples resembling past tasks and interleave them with new data, preventing the network from overwriting old knowledge. The system rehearses synthetic versions of its own history.

The oldest connection is Geoffrey Hinton's wake-sleep algorithm from 1995. It trained a pair of networks in alternating phases. In the wake phase, real data trained the generative model. In the sleep phase, the generative model produced fantasies — Hinton's term — used to train the recognition model. The algorithm was explicitly inspired by the idea that sleep lets a brain learn from its own generated experience.

Google's DeepDream, released in 2015, made the resemblance vivid in the other direction. By amplifying whatever features a vision network detected in an image, it produced the recursive dog faces and eye-studded architecture that looked, to many viewers, uncannily like a psychedelic or dream state.

Perception as controlled hallucination

The deeper point concerns waking perception, not dreams.

Predictive processing theories, developed by Karl Friston, Andy Clark and others, hold that the brain does not passively receive sensory data. It continuously predicts its inputs from an internal generative model and uses incoming signals mainly to correct those predictions. Anil Seth has popularized a phrase for this: perception is a controlled hallucination. The brain generates; the world constrains.

On that view, dreaming is perception with the constraint removed. The generative model keeps running, but sensory error signals are largely cut off, so nothing reins in what it produces. Dreams are what a world model outputs when the world stops answering.

Analysis. Framed that way, a diffusion model at inference and a language model generating without retrieval are in something like the dream condition by default. They produce from priors, and whatever constraint exists comes from the prompt rather than from ongoing contact with reality. Confabulation is not a malfunction of such systems so much as their native mode, which is why grounding techniques — retrieval, tool use, verification against external data — function like sensory input for a sleeping system. They supply the error signal.

What the parallel does not show

None of this implies that generative models dream in the experiential sense. A human dream is something undergone. The structural parallel — generation from noise under learned priors, weakly constrained by input — says nothing about whether there is anything it is like to be the system doing the generating.

What the parallel does show is that the line between perceiving and hallucinating is not the line between real minds and fake ones. Brains blur it constantly. Perception, memory, imagination and dreaming may all run on the same generative machinery under different constraints. If that is right, the useful question about machine hallucination is not why models make things up. It is how any system, biological or artificial, keeps its generative model honest — and that is a question about grounding, not about consciousness.

Machine Dreaming and Perceptual Hallucination

Remote Agency and Distributed Presence

The P-Zombie Dilemma

Your Site Title

email@example.com