Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 96 of the 96 references it cites were found in a registry or an index; 46 of the 86 sources its claims cite were checked for support.

How Adversarial REM Dreams May Facilitate Creativity, and Why We Become Aware of Them

Open access
Deperrois, Nicolas, Petrovici, Mihai, Jordan, Jakob, Huber, Lukas, Senn, Walter · 2024 · Clinical and Translational Neuroscience

Abstract

This perspective reviews a computational approach inspired by artificial intelligence proposing a framework for how REM dreams can contribute to learning and creativity and how NREM dreams may improve robustness of cortical representations.

Overview

CoChat Summary

TL;DR

This perspective article reviews the Perturbed and Adversarial Dream (PAD) model, a computational framework inspired by generative adversarial networks in which REM dreams are generated by a feedback pathway that adversarially tries to fool a discriminator into classifying internally generated sensory activity as real. The authors argue this adversarial process drives the emergence of semantic cortical representations and creative insight, while NREM replay instead improves robustness to sensory perturbations, and they propose that the same discriminator/conductor circuitry explains why we become aware of REM dreams but less so of NREM dreams.

Key findings

  • In the PAD model, REM dreams combine multiple hippocampal memories (e.g., a "dog" and a "car" memory, or a bird and a human yielding a "flying human") via the feedback (FB) pathway; the discriminator (D) is adversarially given the target "real," and the resulting error is backpropagated to make the FB-generated dream more realistic over learning.
  • Simulations reported in the companion work [23] show that including the REM phase yields distinct semantic clusters in PCA projections of high-level representations and much higher linear-classifier accuracy than Wake-only training; the authors state that "the generation of dreams during REM sleep is essential to organizing high-level representations according to the semantics of the sensorium."
  • NREM sleep, modeled as replay of a single hippocampal memory perturbed by randomized occlusions, is reported to have "little or no impact on the acquisition of semantic representations," but instead makes cortical representations more robust to sensory perturbations; REM and NREM are described as complementary.
  • The model predicts that REM dreams facilitate semantic representations and creative insight, but only if dreams balance fantastic and realistic elements; it also proposes content-awareness (from FF→FB teaching signals) versus state-awareness (from the conductor's signal to D) to distinguish wake, NREM, REM, and lucid dreaming.
  • Cited empirical support includes narcolepsy patients (who enter REM directly and have a higher percentage of REM sleep) showing higher creative potential [34], and creative problem solving being boosted on the verge of sleep/N1 [35], by a 60 min nap versus 60 min rest [45], and by eye closure [46].

Methods

  • Perspective/review article: the authors present an accessible overview of the PAD model, which they previously developed and simulated in Deperrois et al. [23], rather than reporting new experiments.
  • The underlying model comprises a cortical feedforward (FF) pathway with a discriminator network (D), a feedback (FB) pathway generating sensory activity from hippocampal memory, and a "conductor" (C) that gates teaching signals across Wake, NREM, and REM phases; it is trained over repeated wake–sleep cycles on natural images.
  • Model evaluation used PCA of the highest FF layer to visualize latent-space clustering and a linear classifier trained on high-level representations; one epoch consists of 780 Wake–NREM–REM cycles sampling 64 images during Wake, with one epoch sampling roughly 5/6 of the CIFAR10 dataset (60,000 images, 10 classes). The paper also proposes future human experiments (category learning with REM-inhibiting pharmacology, mental imagery training, and perturbed-image categorization after NREM awakening).

Limitations

  • The authors state that measuring dream effects is confounded because dreaming co-occurs with sleep, making it hard to disentangle unique effects of dreaming versus sleeping; it is also difficult to attest whether or how a subject is dreaming, and sleep deprivation is experimentally challenging and ethically delicate.
  • The authors note the model does not explicitly address the proposed roles of dreams in emotion and motivation processing.
  • The paper states that the formation of a good representation in the simulations requires 10,000–40,000 Wake–NREM–REM cycles (roughly 8–27 years at 4 cycles/day), which may extend beyond the period during which humans acquire object representations; the hypothesis that REM dreams become more realistic with age "has yet to be tested."
  • Evident from the text: the figures are described as sketches rather than simulation results ("Images represent a sketch, not simulation results"), and the empirical predictions are proposed designs rather than completed studies.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3390/ctn8020021

Cites

96 of 96 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

31 Valid · 5 Not valid · 21 Not checked · 88 Own results and interpretations

66 more claims · 15 with their cited sources checked

Create a free account to see all 145 claims and to verify the 1 reference still open.

See all claims — free account
  • It may state two separate claims as one. 36
  • Probably not a substantial claim. 49

Create a free account to see these claims.

Citations

Cites

Show all 84 references

Research works printed but not linked: 5

  1. Goodfellow, I.J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y. (2014). Advances in Neural Information Processing Systems 27 (NIPS 2014), Neural Information Processing Systems Foundation, Inc. (NeurIPS).
  2. Bengio, S., Wallach, H., Larochelle, H., Grauman, K., Cesa-Bianchi, N., and Garnett, R. (2018). Advances in Neural Information Processing Systems, Curran Associates, Inc.
  3. Olah, G.A. (2002). A Life of Magic Chemistry: Autobiographical Reflections of a Nobel Prize Winner, John Wiley & Sons.
  4. Goodfellow, I., Bengio, Y., and Courville, A. (2016). Deep Learning, MIT Press. Available online: http://www.deeplearningbook.org.
  5. Nielsen, 2001, REM and NREM Sleep Mentation Differences: A Question of Story Structure? Tore, Sleep Hypn., 3, 9

Cited by

See all 3

Metadata

Authors
Deperrois, N., Petrovici, M., Jordan, J., Huber, L., & Senn, W.
Affiliations
University of Bern (CH)
Corresponding author
Senn, Walter
Venue
Clinical and Translational Neuroscience
Publisher
MDPI AG
Published
2024-05-31
Volume
8
Issue
2
Pages
21
ISSN
2514-183X (ISSN-L)
OpenAlex
W4399214191
Cited by
3
References
84
Citations by year
2026: 2 · 2024: 1
FWCI
0.73
Type
journal article
Language
en
Open access
Locations
  • Clinical and Translational Neuroscience (publishedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
  • European Commission (945539)
  • Manfred Stärk Foundation (945539)
Topics
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Sleep and Wakefulness Research
  • Social Sciences › Psychology › Experimental and Cognitive Psychology › Creativity in Education and Neuroscience
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Memory and Neural Mechanisms
Keywords
REM sleep, dreams, creativity, generative adversarial networks, NREM sleep, semantic representation, lucid dreaming, dream classification, learning

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-08, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-08 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: How Adversarial REM Dreams May Facilitate Creativity, and Why We Become Aware of Them. https://app.cochat.ai/papers/10.3390/ctn8020021
The paper
Deperrois, N., Petrovici, M., Jordan, J., Huber, L., & Senn, W. (2024). How Adversarial REM Dreams May Facilitate Creativity, and Why We Become Aware of Them. Clinical and Translational Neuroscience. https://doi.org/10.3390/ctn8020021

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.