Research note 01

The landmark visual-decoding experiment

Horikawa and colleagues used fMRI as participants fell asleep, woke them repeatedly for reports, and identified broad visual categories in their descriptions. Classifiers trained on waking responses to images predicted some reported categories from pre-awakening activity above chance.^[1]

The decoder did not view the dream. It linked distributed activity patterns to labels such as people, buildings or vehicles after extensive participant-specific training.

What the machine actually learns

A decoder begins with paired examples: a measured brain pattern and a label assigned to it. During waking training, a participant may view many images while fMRI records changes in blood oxygenation. The model learns which distributed patterns tend to accompany broad classes of visual features. During sleep, researchers then ask whether a newly measured pattern resembles any of those learned relations.

This procedure explains both the achievement and the limit. If the candidate labels are “person,” “building” and “vehicle,” the model may choose among those categories. It has not learned every object the dreamer could encounter, and it has not recovered colour, dialogue, emotion, movement or narrative continuity. The verbal report used as ground truth is also produced after awakening, when content may already have been forgotten or reorganised.

A later analysis used features from a deep neural network trained for object recognition. Mid- to high-level features decoded from dream fMRI patterns corresponded above chance with the reported object categories.^[2] This suggests that dreamed and perceived objects can share hierarchical visual representations. It does not show that the model reconstructed the exact dreamed object. The analysis matched neural patterns to category-level features drawn from an image database.

Research note 02

Presence, qualities and content

Other work asks simpler questions: was there any experience, was it visual, did it contain movement, or did the dreamer recognise a face? These targets may be more tractable than reconstructing a full narrative.

High-density EEG can relate local frequency patterns to subsequent reports, but reliable generalisation remains difficult. EEG-based models may perform well within a selected dataset and then lose accuracy when tested on a person not represented during training.^{[3], [4]}

The ground-truth problem

Dream decoding has an unusual measurement problem: the experience and the report are not identical. Researchers cannot pause the dream, inspect it independently and compare it with the participant’s description. They must treat a later verbal report as the best available label.

Even a careful immediate report leaves uncertainty. A dreamer may remember a station but forget the train, describe a stranger as a friend because that identity felt obvious, or use the word “house” for a structure that changed throughout the scene. Automated text analysis adds another layer of decisions about synonyms, categories and context.

This is why successful decoding requires strict separation of training and test data, blinded analysis and transparent reporting of all candidate categories. A model should not be evaluated on examples that helped to tune it. It should also be tested on new awakenings and, ideally, new participants. Recent machine-learning studies using public EEG data have reported classification accuracies above 0.85 within their analyses, yet explicitly note poorer generalisation to unseen individuals.^[4] That limitation is central, not a technical footnote.

Research note 03

Why reconstruction headlines overreach

Brain decoding is an inference problem: a model chooses among predefined possibilities using noisy signals and training data. A plausible picture generated from a category is not a recovered image from the dream.

fMRI is slow relative to rapid changes in dream scenes, dream reports are incomplete, and training sets contain assumptions about how words map to images and neural patterns.

The distinction is easy to lose in a striking illustration. If a decoder predicts “building” and a separate image generator produces a detailed castle, only the category prediction came from the neural data. The towers, windows, weather and artistic style were supplied by the generator unless the experiment measured evidence for them.

Research note 04

Future possibilities and ethics

Larger shared datasets, improved sensors and better models may increase prediction of broad features. Interactive lucid-dream experiments could provide tighter timing labels.

Progress also raises questions about consent, neural privacy and sensational claims – anyone fancy a bit of future-crime? Current systems cannot secretly read an ordinary person’s dreams, and it would be misleading to market them as doing so.

The legal history of brain-based evidence is worth examining, including controversial attempts to use brain scans in criminal investigations. Such cases do not validate dream reading and should not be treated as proof that neural data can reveal guilt, intention or private meaning.

From tiny samples to shared data

Dream studies are expensive because each usable observation may require an overnight recording, a precisely timed awakening and a report. Small datasets make it easy for a model to learn individual quirks or laboratory-specific noise rather than a general marker of dreaming.

The DREAM database was created to address part of this problem. Its initial release combined 20 datasets, 505 participants and 2,643 awakenings with at least 20 seconds of pre-awakening M/EEG and standardised report classifications.^[5] Analyses showed that objective EEG features could predict reported conscious experience in both REM and non-REM sleep.

This is a step towards reproducibility, not a library of decoded dream films. The outcome label is commonly experience versus no experience, or another broad report category. Shared data allow independent teams to test whether a marker survives changes in equipment, procedures and participants. They also make negative findings more informative by revealing where a promising classifier fails.

The most credible future progress will probably come in layers: first detecting whether experience is likely, then estimating broad qualities such as visual imagery or movement, and only later attempting more detailed content. Each layer needs its own error rate and validation. Combining them into a polished narrative would conceal uncertainty rather than solve it.

References

1. Horikawa T, Tamaki M, Miyawaki Y, Kamitani Y. Neural decoding of visual imagery during sleep. Science. 2013;340(6132):639-642. doi:10.1126/science.1234330. PMID: 23558170.
2. Horikawa T, Kamitani Y. Hierarchical neural representation of dreamed objects revealed by brain decoding with deep neural network features. Front Comput Neurosci. 2017;11:4. doi:10.3389/fncom.2017.00004. PMID: 28197089.
3. Siclari F, Baird B, Perogamvros L, et al. The neural correlates of dreaming. Nat Neurosci. 2017;20(6):872-878. doi:10.1038/nn.4545. PMID: 28394322.
4. Moctezuma LA, Molinas M. Unlocking dreams and dreamless sleep: machine learning classification with optimal EEG channels. Biomed Res Int. 2025;2025:3585125. doi:10.1155/bmri/3585125. PMID: 39963589.
5. Wong W, Herzog R, Andrade KC, et al. A dream EEG and mentation database. Nat Commun. 2025;16(1):7495. doi:10.1038/s41467-025-61945-1. PMID: 40804039.
6. Mallett R, Konkoly KR, Nielsen T, Carr M, Paller KA. New strategies for the cognitive science of dreaming. Trends Cogn Sci. 2024;28(12):1105-1117. doi:10.1016/j.tics.2024.10.004. PMID: 39500684.