Machine-learning decoders trained on waking stimulus perception can predict the visual contents reported after sleep-onset awakenings from fMRI activity patterns (Horikawa et al., 2013), but that demonstration was confined to NREM stages 1-2 at sleep onset in three highly trained subjects, and decoding based on sleep data alone was much weaker. Whether specific spontaneous dream contents can be decoded during REM sleep and consolidated nocturnal sleep, and whether decoders can be made robust without exclusive reliance on waking stimulus training, have remained open. Deep-neural-network feature decoding (2017) reused the same sleep-onset dataset; high-density EEG work decoded dream presence versus absence - not specific contents - during nocturnal NREM and REM sleep in dozens of subjects; memory-reactivation studies decoded trained, cued material during overnight sleep; and a 2026 contrastive-learning study made wake-trained EEG decoders generalize to overnight sleep in 135 participants, but only for pre-trained semantic categories. No published study is known to decode specific spontaneous dream visual contents outside sleep onset, and no robust sleep-only decoder for such contents is known to exist.

Background

Functional magnetic resonance imaging (fMRI) measures blood-oxygenation-level-dependent (BOLD) signals that indirectly index local neural activity. Multi-voxel pattern analysis (MVPA) fits a statistical 'decoder' that predicts, from the fine-grained pattern of activity across many fMRI voxels, which stimulus a person is viewing; applied to internally generated states, such decoders can also read out mental imagery. Sleep alternates between non-rapid-eye-movement (NREM) stages N1-N3 and rapid-eye-movement (REM) sleep. Dream reports can be obtained after awakening from any stage, with vivid, perception-like dreams most frequent after REM awakenings; the 'sleep-onset' (hypnagogic) period is the transition from wakefulness into N1/N2 at the beginning of a sleep episode. In the serial-awakening paradigm, sleepers are awakened repeatedly and asked to report the last thing they experienced; because the dream itself is private, these retrospective verbal reports are the only source of ground-truth labels for dream contents.

Horikawa, Tamaki, Miyawaki and Kamitani (2013, Science) combined these ingredients: subjects slept in the MRI scanner with simultaneous EEG recording, were awakened at EEG signatures of sleep onset, and reported their visual experience. Words describing reported objects and scenes were mapped onto WordNet synsets - categories of a lexical database that groups semantically similar words - and the fMRI volumes immediately before each awakening were labeled with the reported categories. Decoders trained in separate waking sessions, in which the same subjects viewed object and scene images, were then applied to the sleep data: they classified, detected, and identified dreamed object and scene categories above chance in higher-order visual areas, indicating that specific visual experience during sleep is carried by brain activity patterns shared with stimulus perception. The authors' follow-up (Horikawa and Kamitani, 2017) showed that features of a deep neural network (DNN) - a layered machine-learning model trained for object recognition, used as a proxy for the hierarchy of visual features from simple edges to object-level semantics - could be decoded from the same sleep-onset fMRI data and matched to the dreamed object categories above chance.

These demonstrations, however, rested on decoders trained on waking stimulus perception, on data from sleep-onset NREM stages 1-2 only, and on a tiny cohort of highly trained subjects; the authors reported that decoding based on sleep data itself was far weaker, and they left open whether their approach generalizes to REM sleep and consolidated nocturnal sleep, together with the associated measurement problems (possible imagery contaminating the pre-awakening normalization baseline, movement and report-related activity after awakening, and incomplete category coverage from verbal reports).

Neighboring results frame why these questions are hard and consequential. High-density electroencephalography (hd-EEG) studies of nocturnal sleep (Siclari et al., 2017) found that a 'posterior hot zone' - a region over posterior cortex where low-frequency power decreases and high-frequency power increases - distinguishes dream reports from reports of no experience in both NREM and REM sleep, and that the location of high-frequency activity correlates with broad dream content classes such as faces, spatial settings, movement, and speech; but this presence-level decoding never identified specific dreamed objects or scenes. Other work decoded the reprocessing of recently learned images from overnight sleep EEG (Schönauer et al., 2017) and, most recently, used contrastive learning to make wake-trained EEG decoders generalize to overnight sleep across many participants (Chen et al., 2026) - yet only for trained or cued semantic categories, not spontaneous dream contents. Whether the specific-content decoding achieved at sleep onset can be extended across sleep states and training regimes is the unresolved target of this problem.

Problem Statement

Can the specific visual contents of spontaneous dream experiences be decoded from human brain activity during natural, nocturnal sleep - beyond the sleep-onset NREM stage 1-2 setting of the original 2013 demonstration - and can such decoding be made robust and general?

The single scientific objective is objective measurement of spontaneous dream content across the sleep states in which most dreaming occurs. A complete answer consists of an experimental demonstration (or an adequately powered, rigorously controlled negative result) that addresses, as jointly necessary facets of that objective: (1) decoding of specific dream contents - at least at the level of object or scene categories, or comparably specific contents labeled from post-awakening verbal reports - from brain activity measured before awakenings during REM sleep and/or consolidated nocturnal NREM sleep, at accuracies above chance under cross-validation in which no report contributes to both training and testing; (2) control of the measurement confounds the source study identified, so that decoding is not an artifact of movement, auditory stimulation, or the report process itself (for example through time-course analyses around awakening and appropriate normalization baselines); and (3) characterization of whether decoders trained on sleep data alone, or otherwise robust to the wake-to-sleep distribution shift, can achieve decoding comparable to waking-stimulus-trained models. These facets are jointly necessary because decoding restricted to the sleep-onset period would not constitute objective measurement of dream content across sleep; the problem is not narrowed to any one facet, sleep stage, imaging modality, or decoder class.

References

  1. Horikawa T., Tamaki M., Miyawaki Y., Kamitani Y. (2013). Neural decoding of visual imagery during sleep. Science 340(6132):639-642. DOI: 10.1126/science.1234330. https://doi.org/10.1126/science.1234330
  2. Horikawa T., Kamitani Y. (2017). Hierarchical neural representation of dreamed objects revealed by brain decoding with deep neural network features. Frontiers in Computational Neuroscience 11:4. DOI: 10.3389/fncom.2017.00004 (also arXiv:1611.09520). https://doi.org/10.3389/fncom.2017.00004
  3. Siclari F., Baird B., Perogamvros L., Bernardi G., LaRocque J. J., Riedner B., Boly M., Postle B. R., Tononi G. (2017). The neural correlates of dreaming. Nature Neuroscience 20(6):872-878. DOI: 10.1038/nn.4545. https://doi.org/10.1038/nn.4545
  4. Schönauer M., Alizadeh S., Jamalabadi H., Abraham A., Pawlizki A., Gais S. (2017). Decoding material-specific memory reprocessing during sleep in humans. Nature Communications 8:15404. DOI: 10.1038/ncomms15404. https://doi.org/10.1038/ncomms15404
  5. Wong W., Noreika V., Móró L., Revonsuo A., Windt J., Valli K., Tsuchiya N. (2020). The Dream Catcher experiment: blinded analyses failed to detect markers of dreaming consciousness in EEG spectral power. Neuroscience of Consciousness 2020(1):niaa006. DOI: 10.1093/nc/niaa006. https://doi.org/10.1093/nc/niaa006
  6. Konkoly K. R., Appel K., Chabani E., Mangiaruga A., Gott J., Mallett R., Caughran B., Witkowski S., Whitmore N. W., Mazurek C. Y., Berent J. B., Weber F. D., Türker B., Leu-Semenescu S., Maranci J. B., Pipa G., Arnulf I., Oudiette D., Dresler M., Paller K. A. (2021). Real-time dialogue between experimenters and dreamers during REM sleep. Current Biology 31(7):1417-1427. DOI: 10.1016/j.cub.2021.01.026. https://doi.org/10.1016/j.cub.2021.01.026
  7. Mallett R., Konkoly K. R., Nielsen T., Carr M., Paller K. A. (2024). New strategies for the cognitive science of dreaming. Trends in Cognitive Sciences 28(12):1105-1117. DOI: 10.1016/j.tics.2024.10.004. https://doi.org/10.1016/j.tics.2024.10.004
  8. Moctezuma L. A., Molinas M., Abe T. (2025). Unlocking dreams and dreamless sleep: machine learning classification with optimal EEG channels. BioMed Research International 2025:3585125. DOI: 10.1155/bmri/3585125. https://doi.org/10.1155/bmri/3585125
  9. Moctezuma L. A., Molinas M., Abe T. (2025). From high- to low-density EEG for automatic classification of dream experiences during stage 2 of NREM. Sleep Advances 6(4):zpaf066. DOI: 10.1093/sleepadvances/zpaf066. https://doi.org/10.1093/sleepadvances/zpaf066
  10. Horikawa T. (2025). Mind captioning: evolving descriptive text of mental content from human brain activity. Science Advances 11(45):adw1464. DOI: 10.1126/sciadv.adw1464. https://doi.org/10.1126/sciadv.adw1464
  11. Chen Z., Zheng H., Zhou J., Zheng L., Lin P., Wang H., Busche M. A., Behrens T., Dolan R., Liu Y. (2026). Interpreting human sleep activity through neural contrastive learning. Neuron (published online April 2026). DOI: 10.1016/j.neuron.2026.03.028. https://doi.org/10.1016/j.neuron.2026.03.028