The overfitted brain: Dreams evolved to assist generalization

Dreaming remains a mystery to neuroscience. While various hypotheses of why brains evolved nightly dreaming have been put forward, many these are contradicted by the sparse, hallucinatory, and narrative nature dreams, that seems lack any particular function. Recently, research on artificial neural networks has shown during learning, such face ubiquitous problem: overfitting dataset, which leads failures in generalization therefore performance novel datasets. Notably, techniques researchers employ rescue overfitted generally involve sampling from an out-of-distribution or randomized dataset. The brain hypothesis is organisms similarly challenge fitting too well their daily distribution stimuli, causing poor generalization. By hallucinating sensory stimulation every night, able generalizability its perceptual cognitive abilities increase task performance. Understanding biological function sleep advanced considerably past decade. However, no equivalent understanding dreams emerged. Contemporary neuroscientific theories often view as epiphenomena, proposals for phenomenology themselves. Now, recent advent deep (DNNs) finally provided conceptual framework within understand dreams. all DNNs issue they learn, when one dataset increases but network's fails generalize (often measured divergence training versus testing datasets). This problem solved modelers via “noise injections” form noisy corrupted inputs. goal this paper argue faces similar combat brain's learning. That is, mechanism increasing creation inputs stochastic activity across hierarchy structures. Sleep loss, specifically dream can still memorize learn appropriately. Herein ”overfitted hypothesis” explicitly developed then compared contrasted with existing contemporary Existing evidence surveyed both neuroscience set testable predictions forward be pursued vivo silico. During Covid-19 pandemic 2020, those isolation reported vividness frequency dreams,1Weaver C. Why Am I Having Weird Dreams Lately?. New York Times, 2020https://www.nytimes.com/2020/04/13/style/why-weird-dreams-coronavirus.htmlGoogle Scholar even leading #pandemicdreams trend Twitter. Yet so little understood there only speculative answers behind widespread change behavior. despite fact humans spend hours night deprivation highly damaging animals2Albert I. Cicala G.A. Siegel J. behavioral effects rem rats.Psychophysiology. 1970; 6: 550-560Crossref PubMed Scopus (101) Google Scholar; indeed, homeostatically regulated appears ”dream drive.”3Dement W. effect deprivation.Science. 1960; 131: 1705-1707Crossref (279) In addition, conserved species,4Siegel J.M. Evolution REM Sleep. CRC Press, 1999Google indicating essential purpose.5Cirelli Tononi G. Is essential?.PLoS Biol. 2008; 6Crossref (348) finding themselves evaded resolution. ”null theory” byproducts other adaptions,6Domhoff G.W. neurocognitive theory at age 20: assessment comparison four dreaming.Dreaming. 2019; 29: 265Crossref (1) ”spandrels,”7Flanagan O.J. Souls: Sleep, Dreams, Conscious Mind. Oxford University 2000Google adaptive prevent moving periods not adapted to, ensuring diurnal animals inactive night,8Siegel animals: state inactivity.Principles Pract. Med. 2011; 5: 126-138Crossref (3) Scholar,9Siegel Rem sleep: psychological paradox.Sleep Rev. 15: 139Crossref (51) taken seriously. It notable approximately 50% 70% time subjects report having upon sudden waking sleep, more being later night.10Foulkes W.D. Dream reports different stages sleep.J. Abnormal Social Psychol. 1962; 65: 14Crossref (0) Scholar,11Stickgold R. Malia A. Fosse Propper Hobson J.A. Brain-mind states: longitudinal field study sleep/wake factors influencing mentation length.Sleep. 2001; 24: 171-179Crossref Reports who keep regular journals, each morning recollecting indicate preserving memory significantly practice.12Robb We Dream: Transformative Power Our Nightly Journey. Houghton Mifflin Harcourt, 2018Google All hints individuals may regularly underestimate how much actually spending sleep. experience.13Windt Dreaming: A Conceptual Framework Philosophy Mind Empirical Research. MIT press, 2015Crossref few appropriately account particularly contents. What purpose strange state? overall some relationship learning was known Roman orator Quintilian.14Duff K. Secret Life Simon Schuster, 2014Google Yet, discussed detail next section, relate storage, replay, emotional processing, epiphenomena. viable about stands contrast physiology stages.15Lee S.H. Dan Y. Neuromodulation states.Neuron. 2012; 76: 209-222Abstract Full Text PDF (296) As originally discovered lesion studies supported genetic knockout studies, brought far-reaching subcortical neuromodulatory systems, system necessary, redundancy sustained.16Jones B.E. From sleeping: neuronal chemical substrates.Trends Pharmacol. Sci. 2005; 26: 578-586Abstract (439) general, multifaceted arousal excitatory wake it greatest firing, becomes quiescent bring although true especially establish rapid eye movement (REM). classic signature slow waves, waves traverse cortex, identified dominate electroencephalogram (EEG) less than 1 Hz. state, cortex become bistable, oscillating between intense firing referred “down states” wherein neurons silent. should noted spectrum sometimes region experiencing wave synchronized regions.17Nir Staba R.J. Andrillon T. Vyazovskiy V.V. Cirelli Fried Regional spindles human sleep.Neuron. 70: 153-169Abstract (503) itself traditionally broken down into non-REM (NREM) associated NREM occurs throughout stages,18Oudiette D. Dealberto M.J. Uguccioni Golmard J.L. Merino-Andreu M. Tafti Garma L. Schwartz S. Arnulf without sleep.Conscious. Cogn. 21: 1129-1140Crossref (43) rarest “deepest” stage NREM, 3, surface EEG reflects low-frequency cortical waves. Recent neuroimaging sudden-waking experiments demonstrated result localized wake-like firing.19Siclari F. Baird B. Perogamvros Bernardi LaRocque J.J. Riedner Boly Postle B.R. correlates dreaming.Nat. Neurosci. 2017; 872Crossref (225) On average, high-frequency signals posterior areas were most correlated waking. Despite ubiquity case strongly dreaming, 80% 90% awakening again involves difficult find does contain all.20Cipolli Ferrara De Gennaro Plazzi Beyond neuropsychology dreaming: insights basis new recording analysis.Sleep 35: 8-20Crossref (8) Although early present “thought-like” simple, REM, incredibly complex fully structure.21Siclari dreaming.The Neurol. conciousness. 2016; : 107-128Google Therefore, while story occurring solely complicated, associated. far currently known, brain-wide where single event, activation default-mode network dreaming.22Domhoff substrate subsystem default network?.Conscious. 1163-1174Crossref (61) Consider, instance, higher frontal lobe predicted affect subject's dream.23Sikka P. Revonsuo Noreika V. Valli Eeg alpha asymmetry affect: oscillations over right presleep wakefulness predict anger dreams.J. 39: 4775-4784Crossref (7) Or consider lucid movements sensorimotor cortex.24Dresler Koch S.P. Wehrle Spoormaker V.I. Holsboer Steiger Sämann P.G. Obrig H. Czisch Dreamed elicits cortex.Curr. 1833-1837Abstract (97) nuances redundancies systems play, makes sense dreamless regulates level consciousness, acetylcholine-containing REM. Neuromodulatory also create conditions muscle atonia acted out body dangerous parasomnia called behavior disorder.25Schenck C.H. Bundlie S.R. Mahowald M.W. disorder: clinical, developmental, perspectives 16 years after formal identification sleep.Sleep. 2002; 25:120–138; Crossref (584) sleep? distinct physiological states neuromodulation suggests answering requires identifying multiple functions, dreaming. Across Tree Life, whole conserved; mammals somewhere 4 20 h sleeping.26Joiner W.J. Unraveling evolutionary determinants sleep.Curr. R1073-R1087Abstract (75) There Caenorhabditis elegans sleeps.27Trojanowski N.F. Raizen D.M. Call worm sleep.Trends Neurosciences. 54-62Abstract 2 decades significant progress comes whole, First, discovery led clear least aspect glymphatic system, showing flushing metabolites cerebral spinal fluid.28Xie Kang Xu Q. Chen Liao Thiyagarajan O’Donnell Christensen D.J. Nicholson Iliff et al.Sleep drives metabolite clearance adult brain.science. 2013; 342: 373-377Crossref (1959) activity, had waste partly sleep's restorative aspect. Glymphatic low high under certain types anesthesia. anesthesia, amount delta power dominates EEG,29Hablitz L.M. Vinitsky H.S. Sun Stæger F.F. Sigurdsson Mortensen K.N. Lilius T.O. Nedergaard Increased influx eeg heart rate mice anesthesia.Sci. Adv. eaav5447Crossref (89) anticorrelated established. occur events; unknown if might depending background neurophysiological conditions. Another important Synaptic Homeostasis Hypothesis (SHY).30Tononi synaptic homeostasis: hypothesis.Brain Res. Bull. 2003; 62: 143-150Crossref (734) According SHY, net potentiation brain, which, left unchecked, would lead saturation weight cessation learning.31Tononi homeostasis.Sleep 2006; 10: 49-62Crossref (1297) SHY hypothesizes trigger down-scaling weights. indiscriminate ensures relative weights synapses kept proportional removing risk saturation. served model generated number empirical findings.32Bushey structural Drosophila.Science. 332: 1576-1581Crossref (230) At same time, triggered fecund debate investigation.33Frank M.G. Erasing shy?.Neural Plast. 2012: 264378Crossref (81) Traditionally, frequencies explored global could easily fail pattern same, since nonlinear mechanisms. Even rectified linear unit (ReLU) functions likely scaling dropped below zero; sigmoid impact function, way. indicates two parts: metabolic cellular repair occurs, contribution improvements tasks dreaming.34Cao Herman A.B. West G.B. Poe Savage V.M. we quantitative analysis reveals abrupt transition reorganization development.Sci. 2020; eaba0398Crossref (5) Ultimately, phase lacks explicit sleep.35Walker Sleep: Science Dreams. Penguin UK, 2017Google order offer outlines idea constantly danger overfitting, based help mitigate issue. brian (OBH). formally fill-out OBH investigating fits composed statistically experiences organism, improve generality robustness animal's representations, cognition, generating data outside organism's ”training set” warped way.36Hoel E. Enter supersensorium: art netflix?.Baffler. https://thebaffler.com/salvos/enter-the-supersensorium-hoel.htmlGoogle stochasticity (such sparse inputs) critical machine learning.37Sabuncu M.R. Intelligence plays dice: learning.arXiv. will discussed, data, matches better theories, draws antecedents approaches specific roof, additionally roots common practices unique tested computational modeling vivo. must outline aspects explain themselves, experience. Specifically, phenomenology. Consider three phenomenological properties sparseness vivid life information (i.e., detail). universal, examples include blurring text impossibility reading, using phones, calculations.38Hartmann do 3 r’s: implications mentation.Dreaming. 2000; 103-110Crossref Second, hallucinatory quality unusual way simple repetition events memories). includes concepts exist normally strict categories (a person another person, house spaceship, on). Third, property sequences ordered narrative, albeit fabulist one. see, posits spite serve purpose, because them. section explores existent supporting (or thereof) integrate with, explain, health descendant Freudian psycholanalysis.39Freud Cronin A.J. Interpretation Read Books Ltd, 2013Google Freud's expressions taboo frustration discredited, historical association of, for, processing. regulation like somehow fear extinction.40Levin Nielsen Nightmares, bad emotion dysregulation: review dreaming.Curr. Dir. 2009; 18: 84-88Crossref Such reason act somewhat therapy treatment phobias, provide safe space “rehearsals” toward fearful things make them frightful.41Scarpelli Bartolacci D'Atri Gorgoni functional role processes.Front. 459Crossref (23) fears nightmares kind irrational faced nor nightmarish general attenuated, evolutionarily quite useful. “emotional thermostat” regulate emotions.42Cartwright mood system.in: Principles Practice Medicine. Elsevier, 2005: 565-572Crossref Scholar,43Walker M.P. van Der Helm Overnight therapy? processing.Psychol. 135: 731Crossref (551) journals recovery disorders depression,44Cartwright Agargun M.Y. Kirkby Friedman J.K. Relation concerns.Psychiatry 141: 261-270Crossref (88) sample size sort prohibitively small oneirology. processing centers amygdala show greater wake,45Hobson Pace-Schott E.F. Stickgold Kahn To dream? relevant electrophysiological studies.Curr. Opin. Neurobiol. 1998; 8: 239-244Crossref ranges widely emotions rewards motivations.46Janak P.H. Tye K.M. circuits behaviour amygdala.Nature. 2015; 517: 284-292Crossref (795) changes disorders.47Kupfer latency: psychobiologic marker primary depressive disease.Biol. Psychiatry. 1976; 11: 159PubMed changed reduced disorders,48Gillin J.C. Duncan Pettigrew K.D. Frankel B.L. Snyder Successful separation depressed, normal, insomniac data.Arch. Gen. 1979; 36: 85-90Crossref problems general.49Peterson Benca R.M. disorders.Psychiatr. Clin. 1009-1032Google appear issues inhibition irritability.50Gruber Cassoff interplay regulation: future directions.Curr. Psychiatry Rep. 2014; 16: 500Crossref (24) But appropriate holds processes following deprivation, including executive regulation.51Killgore W.D.S. Effects cognition.Prog. Brain 2010; 185: 105-129Crossref (564) Overall, resolving conflicts overwhelming evidence. emphatically emotional. Indeed, emotionally neutral common. Overall slight bias negative affect,52Merritt Williams Emotion profiles men women.Conscious. 1994; 3: 46-60Crossref simply valence memorable (and worth noting studying joy/elation anxiety/fear). Given whatever emotions, strong regulation. Perhaps consolidation storage,53Diekelmann Born sleep.Nat. 114-126Crossref (29) proposed replay.54Wamsley E.J. offline consolidation.Curr. 14: 433Crossref dominant metaphor computer: memories need “stored” storing computer file hard drive, storage process. viewpoint held traditional neuroscience, “store” veridically possible,55Marcus Kluge: Haphazard Human 2009Google growing recognition veridical “computer-like” desirable forgetting just important.56Richards B.A. Frankland P.W. persistence transience memory.Neuron. 94: 1071-1084Abstract (110) hypothesis, alternatively accessing previously stored memories, strengthen them, byproduct integrating older ones. line theory, full literature beyond scope paper.57Vorster A.P. mammals, birds invertebrates.Neurosci. Biobehavioral 50: 103-119Crossref debate. Specifically very broad rarely meant s