Every night, something extraordinary happens to the human mind. We close our eyes in one world and, for a few hours, inhabit another. In that world, we may speak with people who died years ago, revisit places that never existed, walk through impossible landscapes or witness events that violate every law of physics we know. Yet while these experiences are unfolding, they rarely seem strange. We accept them with remarkable ease. We respond emotionally, make decisions, experience fear, joy, embarrassment and curiosity, all while remaining largely unaware that the reality surrounding us has been generated entirely from within.
The true mystery of dreaming is not that dreams are bizarre. The mystery is that they feel real.
This observation may seem trivial, but it lies at the heart of one of the most profound questions in neuroscience. How can the brain generate a convincing reality in the absence of the external world? During sleep, there is no city beyond our window, no conversation taking place in the room, no mountain rising on the horizon. Yet the subjective experience remains vivid, coherent and immersive. The sleeping brain somehow creates a world convincing enough to replace physical reality for hours at a time.
For much of human history, dreams were interpreted through mythology, religion or philosophy. Ancient cultures saw them as messages from gods, warnings from ancestors or glimpses into hidden dimensions of existence. Later, psychologists attempted to understand dreams as symbolic expressions of unconscious desires and conflicts. Although these interpretations differed dramatically, they shared a common assumption: dreams were meaningful because they revealed something concealed beneath ordinary awareness.
Modern neuroscience has taken a different approach. Rather than asking what dreams mean, researchers increasingly ask how dreams are generated. This shift in perspective may seem subtle, but it has transformed the study of sleep. Instead of treating dreams as mysterious messages waiting to be decoded, scientists have begun examining them as products of a biological system. The question is no longer what dreams are trying to tell us. The question is what dreams reveal about the machinery that creates conscious experience itself.
One of the most influential modern definitions of dreaming comes from philosopher and dream researcher Jennifer Windt, who describes dreams as immersive spatiotemporal hallucinations. At first glance, the word hallucination may seem provocative. In everyday language, it is strongly associated with psychiatric illness or neurological dysfunction. Yet in its technical sense, the term is surprisingly precise. A hallucination is a perception that occurs in the absence of the corresponding external stimulus. A person sees something that is not present, hears something that is not occurring or experiences events that do not exist in the surrounding environment.
By this definition, dreaming qualifies remarkably well.
The difference lies in scale. Most hallucinations occur within an already existing reality. A person might hear a voice in an otherwise ordinary room or see an object that others cannot perceive. Dreams go much further. Rather than adding elements to reality, they temporarily replace reality altogether. During sleep, the brain constructs space, time, people, objects, narratives and even the sense of being a self moving through those events. It generates an entire experiential world and then presents that world as reality.
This observation leads to an important conclusion. The brain possesses an extraordinary capacity to generate coherent experiences from internally available information alone. The world of a dream may not correspond to physical reality, but it is nonetheless a reality in the phenomenological sense. It is experienced. It is inhabited. It feels real.
The more researchers studied dreams, the more they realised that dream content is not simply a replay of waking life. Dreams rarely resemble recordings of the previous day. Instead, they appear to be assembled from fragments of memory, emotion, expectation, bodily sensation and imagination. Familiar elements are reorganised into unfamiliar combinations. A childhood home may merge with a modern office building. A deceased relative may appear in a future event. A city never visited can somehow feel intimately familiar.
The brain is not reproducing reality.
It is generating possibilities.
This insight became particularly important during the second half of the twentieth century, when neuroscience began moving away from passive models of perception. Traditionally, perception was imagined as a relatively straightforward process. Information entered through the senses, travelled into the brain and was transformed into conscious experience. The brain functioned as a sophisticated receiver, analysing signals arriving from the outside world.
Over the last several decades, however, a very different picture has emerged. Increasingly, researchers have come to view the brain not as a passive receiver of information but as an active generator of models.
This perspective is most closely associated with predictive processing theories, particularly the work of neuroscientist Karl Friston. Although the mathematical details of these theories are complex, the central idea is surprisingly intuitive. The brain does not wait passively for sensory information to arrive before constructing a representation of the world. Instead, it continuously generates predictions about what is likely to be present in the environment. Sensory information then serves primarily to confirm, modify or correct those predictions.
From this perspective, perception resembles hypothesis testing more than signal reception.
The brain constantly asks itself what is most likely happening. What object is producing a particular visual pattern? What sound is generating a particular vibration in the ear? What event is unfolding in the surrounding environment? Incoming sensory information provides evidence, but the interpretation of that evidence depends heavily on predictions that already exist before the information arrives.
Examples of this process appear everywhere in everyday life. We can read sentences containing spelling mistakes because the brain predicts intended words. We can understand speech in noisy environments because the brain predicts likely sounds. We can navigate familiar places without consciously analysing every detail because the brain predicts what should be present before we look.
Prediction is not a rare feature of cognition.
It may be one of its fundamental principles.
Seen from this perspective, dreaming becomes especially fascinating. During wakefulness, the brain's internal models are continuously constrained by sensory evidence. The external world functions as a correction mechanism, preventing predictions from drifting too far from reality. During sleep, however, much of this corrective input is dramatically reduced. The brain continues generating models, narratives and expectations, but receives far less information with which to test them.
Dreaming may therefore reveal what happens when prediction continues, but correction weakens.
The result is not random chaos. The result is a coherent, immersive reality generated almost entirely from within.
At this point, dreams cease to be merely a curiosity of sleep science. They become a window into a much broader question: how does the brain construct reality in the first place?
The implications become even more striking when we consider what happens outside of sleep. If waking perception depends on a balance between internally generated models and externally supplied evidence, then dreams represent only one point along a much larger spectrum. At one extreme lies the tightly constrained reality of ordinary perception. At another lies the internally generated reality of dreaming. Between these states exist imagination, daydreaming, hallucinations, delusions and psychosis.
Each of these phenomena may reveal something about the same underlying architecture.
Each may reflect a different relationship between prediction and correction.
And it is precisely here that the science of dreaming begins to intersect with one of the most important questions in psychiatry and cognitive neuroscience.
What happens when internally generated models begin to overpower external evidence?
Psychosis and the predictive brain
If dreams reveal what happens when the brain continues generating models while sensory correction is reduced, an unsettling question naturally follows. What happens when something similar begins to occur during wakefulness?
For much of modern history, psychosis was described primarily through its symptoms. Clinicians documented hallucinations, delusions, disorganised thinking and disturbances of perception. These descriptions were invaluable for diagnosis and treatment, but they did not necessarily explain why such experiences occurred. They identified the visible consequences of a process without fully revealing the mechanism behind it.
The emergence of predictive processing theories introduced a different perspective. Rather than asking only what people with psychosis experience, researchers began asking how the brain might arrive at those experiences in the first place. If perception depends on the continuous interaction between prediction and sensory evidence, then disturbances in that interaction could potentially alter reality itself.
This idea necessitates a shift in how we perceive the world. In everyday life, we tend to assume that reality is simply received. Light enters the eyes, sound enters the ears, information reaches the brain, and perception follows. Predictive theories suggest a more dynamic process. The brain is constantly generating hypotheses about what is happening in the world and then comparing those hypotheses against incoming sensory information. Reality, as we experience it, emerges from this ongoing negotiation between expectation and evidence.
Under normal circumstances, the process is remarkably stable. Predictions help us navigate an uncertain environment efficiently. They allow us to recognise familiar faces in poor lighting, understand speech in noisy rooms and anticipate the movements of objects before they occur. Without prediction, perception would be painfully slow and computationally expensive. The brain would have to interpret every sensory detail from scratch.
Yet prediction carries an inherent risk.
A system that relies on expectations must constantly decide how much confidence to place in those expectations and how much confidence to place in incoming evidence. If the balance shifts too far in one direction, perception may become unstable.
Imagine walking alone through a forest at dusk. A brief movement appears in the shadows. It could be a predator, another person or simply wind moving through vegetation. The sensory information is ambiguous. The brain must decide how seriously to take it. If the potential threat is considered highly probable, the nervous system may respond immediately. If the signal is judged unreliable, it may be ignored.
Most of the time these judgments occur automatically and without conscious awareness. Yet they reveal an important feature of perception. The brain is not simply determining what is present. It is continuously estimating how much trust should be assigned to different sources of information.
Predictive processing researchers often describe this in terms of precision. Precision refers to the estimated reliability of a signal. A highly precise signal deserves attention and can drive learning. A low-precision signal may be treated as noise and ignored.
This concept becomes especially important when considering psychosis.
One influential interpretation proposes that psychotic experiences may arise when the brain begins assigning unusual levels of precision to signals that would normally be considered insignificant. Random coincidences start to feel meaningful. Neutral events acquire personal significance. Weak sensory impressions begin to demand interpretation. The brain becomes increasingly sensitive to information that might otherwise have been dismissed.
At first glance, this may sound like a purely perceptual problem. In reality, the consequences extend much further. Human beings are meaning-making organisms. We do not simply perceive events; we explain them. Whenever the brain encounters something unexpected, it attempts to construct a narrative that makes sense of the anomaly.
Imagine noticing a strange sound in your house at night. The sound itself is only a sensory event. Almost immediately, however, the mind begins generating explanations. Was it the wind? A neighbour? An animal? A problem with the building? Perception and interpretation are inseparable.
Now imagine a system in which unexpected events occur more frequently because weak signals are being treated as highly significant. The demand for explanation increases dramatically. The brain begins searching for patterns, causes and narratives capable of making sense of an increasingly unpredictable world.
This may help explain why hallucinations and delusions often appear together.
Hallucinations provide unusual experiences.
Delusions provide explanations for those experiences.
The relationship is not always straightforward, but the connection is understandable. A mind confronted with persistent anomalies will naturally attempt to build a coherent model capable of accommodating them.
From this perspective, psychosis can be viewed not as a failure of meaning-making but as an extreme expression of it. The brain continues doing what it has always done: constructing reality, generating explanations and attempting to reduce uncertainty. The difference lies in how evidence is weighted and interpreted.
This interpretation does not reduce psychosis to a simple computational error. Human experiences are far more complex than any single theory can capture. Social factors, genetics, developmental history, trauma, neurochemistry and environment all play critical roles. Nevertheless, predictive processing provides a useful framework because it shifts attention from isolated symptoms to the deeper mechanisms that generate experience itself.
The connection to dreaming now becomes difficult to ignore.
Dreams and psychosis are obviously different phenomena. They involve different neurochemical states, different patterns of brain activity and profoundly different consequences for behaviour and wellbeing. Yet both reveal something important about the architecture of consciousness. In each case, the brain demonstrates its capacity to generate a convincing reality from within. The distinction lies not in whether reality is being constructed, but in how tightly that construction is constrained by external evidence.
Dreaming shows what happens when sensory correction is dramatically reduced. Psychosis may reveal what happens when the balance between prediction and evidence becomes disrupted during wakefulness. Both phenomena expose a feature of the mind that usually remains hidden behind the apparent stability of everyday perception.
The implications extend beyond psychiatry. They force us to reconsider what perception actually is. The traditional distinction between reality and imagination begins to appear less absolute than it once seemed. The brain is always constructing models. It is always generating expectations. It is always attempting to explain the causes of its sensory inputs. What changes from one state to another is not the existence of those models but the degree to which they are constrained by evidence.
Seen in this light, the most surprising lesson of psychosis research may not concern psychosis at all. Instead, it concerns ordinary perception. The fact that most people experience a stable and shared reality is not evidence that the brain passively mirrors the world. It may be evidence that the brain possesses extraordinarily effective mechanisms for keeping its internal models aligned with external constraints.
This realization leads directly to another puzzle. If the brain is constantly updating its model of reality, constantly learning from experience and constantly modifying the connections that support those predictions, how does it avoid becoming overwhelmed by its own learning? How does a system that continuously changes itself maintain stability over time?
To answer that question, we must leave the clinic and return to the sleeping brain. Because the next chapter of this story is not about psychosis. It is about learning itself—and the surprisingly high biological price that learning may require.
The cost of learning: Tononi and Cirelli's radical question
At first glance, learning seems like an unqualified good.
From early childhood onward, nearly every measure of cognitive success is tied to our ability to acquire new information, adapt to changing circumstances and modify our behaviour in response to experience. The brain's extraordinary capacity for learning is often celebrated as one of the defining features of human intelligence. We learn languages, recognise faces, master skills, navigate social relationships and build increasingly sophisticated models of the world. In neuroscience, this capacity is known as plasticity—the ability of neural circuits to change as a result of experience.
For decades, plasticity was viewed primarily as a solution. It explained how memories form, how skills improve and how experience shapes the nervous system throughout life.
But eventually a different question began to emerge.
What if plasticity is also a problem?
This was the question that Giulio Tononi and Chiara Cirelli began asking in the early 2000s. Their answer would eventually develop into one of the most influential theories of sleep in modern neuroscience: the Synaptic Homeostasis Hypothesis, often abbreviated as SHY.
The idea began with a simple observation. Learning is not free.
Every meaningful experience leaves a trace in the brain. Every new skill, every successful prediction, every mistake and every emotionally significant event alters patterns of connectivity within neural networks. Although these changes are essential for adaptation, they also carry biological costs. Stronger synapses consume more energy, require more molecular maintenance and occupy more physical space. A brain that continuously strengthens connections cannot do so indefinitely.
This may sound obvious, yet its implications are surprisingly profound.
Imagine a city in which every road that is ever used becomes slightly wider. Every successful route receives additional lanes. Every path that helps people reach a destination becomes permanently reinforced. At first, the system would improve. Traffic would flow more efficiently. Important routes would become easier to navigate.
But what happens after years of continuous expansion?
Eventually, the city would encounter a problem. Roads would begin competing for space. Maintenance costs would increase. Distinguishing major transportation corridors from minor side streets would become increasingly difficult. What began as an adaptive process would eventually threaten the efficiency of the entire system.
Tononi and Cirelli argued that something similar may happen inside the brain.
Throughout wakefulness, countless synapses undergo strengthening as a consequence of learning. The exact mechanisms differ across neural systems, but the overall trend is remarkably consistent. Experience tends to increase synaptic efficacy. Networks become more responsive to patterns that have proven important in the past. The brain continuously modifies itself in order to better predict and navigate the future.
This process is extraordinarily useful.
It is also potentially unsustainable.
The central insight of the Synaptic Homeostasis Hypothesis is that the problem is not memory itself. The problem is accumulation.
If learning continuously increases synaptic strength across large portions of the brain, then a mechanism must exist to prevent those increases from growing without limit. Otherwise, neural networks would gradually become saturated by their own history.
Sleep, according to Tononi and Cirelli, may be that mechanism.
Importantly, the hypothesis does not propose that sleep erases memories. In fact, the theory was developed partly to explain why sleep often improves learning and memory performance. The key idea is subtler than forgetting. During sleep, the brain may reduce overall synaptic strength while preserving the relative importance of significant connections.
An analogy from photography helps illustrate the concept.
Imagine gradually increasing the brightness of every pixel in an image throughout the day. At first, important features become easier to see. Eventually, however, the entire image becomes overexposed. Contrast disappears because everything is bright. The problem is no longer insufficient signal. The problem is too much signal.
Restoring the image does not require deleting information. It requires reducing brightness across the board while preserving differences between strong and weak features.
According to SHY, sleep may perform a similar operation on neural networks.
The goal is not to remove learning.
The goal is to restore contrast.
Strong, important and repeatedly reinforced connections remain relatively strong. Weak, transient and poorly supported changes lose influence. The result is a system that retains what matters while preventing overall activity from drifting toward instability.
When Tononi and Cirelli first proposed this idea, it was elegant but highly speculative. A compelling theory is not enough in science. The challenge was finding evidence that the brain actually behaves this way.
Over the following two decades, researchers began accumulating precisely that evidence.
One of the earliest breakthroughs came from studies examining molecular markers associated with synaptic potentiation. If wakefulness truly produces a net increase in synaptic strength, then measurable biological signatures of strengthening should be more prominent after extended periods of wakefulness than after sleep.
This is exactly what several studies appeared to find.
Markers associated with synaptic potentiation increased during wakefulness and decreased following sleep. Electrophysiological measurements revealed similar patterns. Neural responses often appeared stronger after prolonged waking and reduced after periods of sleep. These findings did not prove the hypothesis, but they provided some of the first evidence that sleep and wakefulness might indeed exert opposite effects on overall synaptic strength.
The story became even more intriguing when researchers began examining the problem at larger scales.
Not all parts of the brain work equally hard during the day. Some networks may be heavily engaged by learning, while others remain relatively inactive. If sleep serves a restorative role, then its effects might be expected to appear locally rather than uniformly across the entire brain.
Evidence supporting this possibility emerged from studies of local sleep. Researchers found that brain regions intensely involved in learning tasks during wakefulness often displayed stronger sleep-related activity later that night. In other words, sleep did not appear to be a completely global process imposed equally on every neural circuit. Instead, specific networks seemed to exhibit signs of greater sleep pressure depending on how heavily they had been used.
This observation hinted at something remarkable.
The sleeping brain may not simply be resting.
It may be selectively recalibrating itself.
Further support came from structural studies examining synapses directly. In some experiments, researchers observed that many synaptic connections became smaller after sleep, particularly weaker and intermediate-strength connections. Importantly, the strongest synapses tended to be preserved. This pattern aligned closely with one of the central predictions of SHY. Sleep appeared not to be indiscriminately weakening all connections but preferentially reducing those most likely to represent noise, transient learning or less important traces of experience.
At this point, the theory began to attract attention beyond sleep research.
Because the problem Tononi and Cirelli were describing sounded surprisingly familiar.
In fact, it sounded almost identical to a problem that computer scientists and artificial intelligence researchers had been struggling with for decades.
The problem was overfitting.
A learning system that strengthens every pattern indiscriminately eventually becomes trapped by its own experience. It begins treating noise as signal, coincidence as structure and transient events as enduring truths. Such a system may appear highly knowledgeable, yet paradoxically become worse at learning.
The solution, both in machine learning and potentially in biology, is not to stop learning.
The solution is to prevent learning from accumulating without limit.
This observation creates an unexpected bridge between sleep research and predictive theories of perception. Because once the brain is understood as a system that continuously updates its internal model of reality, the importance of sleep begins to look very different.
Sleep is no longer simply a period of rest.
It becomes part of the learning process itself.
The brain spends the day modifying its model of the world. During the night, it may perform the equally important task of deciding which modifications deserve to remain. Learning, in this view, is not something that ends when we fall asleep. It continues throughout the night, as neural networks reorganise themselves in preparation for another day of prediction, adaptation and experience.
This realisation sets the stage for a deeper synthesis. Because the problem Tononi and Cirelli describe—the need to prevent neural systems from becoming overwhelmed by their own learning—appears remarkably similar to a problem that sits at the centre of Karl Friston's predictive framework.
Both theories are ultimately concerned with the same question.
How can a system remain flexible enough to learn while remaining stable enough to function?
The answer may lie in understanding sleep not merely as recovery, but as a process of biological regularization—a nightly recalibration of the machinery that constructs reality itself.
Sleep as a biological regularisation
By the time we arrive at this point in the story, sleep begins to look very different from the way it is usually portrayed. For decades, popular discussions of sleep have focused on restoration. Sleep was described as a period during which the brain recovered from the demands of wakefulness, replenished depleted resources and consolidated memories. None of these functions is necessarily wrong, but they may describe only part of a much larger process.
If the Synaptic Homeostasis Hypothesis is even partially correct, then sleep is not simply a recovery period. It is an active phase of network management. The sleeping brain is not shutting down; it is reorganizing itself. More importantly, it may be solving a fundamental problem faced by every learning system, whether biological or artificial: how to continue learning without becoming trapped by its own history.
This problem is familiar to anyone who works with machine learning. A neural network trained on real-world data must constantly balance two competing demands. On the one hand, it needs to learn from experience. If it fails to incorporate new information, it becomes rigid and incapable of adaptation. On the other hand, it cannot treat every piece of information as equally important. If it does, the model begins memorising noise, random fluctuations and accidental patterns that have little predictive value. Eventually, it becomes less capable of understanding the world precisely because it has learned too much.
Engineers refer to this problem as overfitting. A model becomes excessively tuned to specific experiences and loses its ability to generalise. Instead of capturing the underlying structure, it becomes trapped by details. The result is a system that performs impressively in familiar situations but struggles when confronted with novelty.
Biological brains face a remarkably similar challenge.
Every day, billions of neurons participate in an immense process of adaptation. The brain continuously updates expectations, strengthens associations, encodes memories and modifies behavioural strategies. Most of these changes are useful. Some are extremely important. Many are trivial. A few are simply noise.
The difficulty lies in distinguishing among them.
At any given moment, the brain cannot know with certainty which experiences will prove significant in the future. A brief conversation may alter the course of a career. A seemingly insignificant event may later become critically important. Conversely, many emotionally intense experiences may ultimately have little lasting relevance.
The nervous system, therefore, faces a difficult computational problem. It must remain sufficiently plastic to learn from experience while avoiding the accumulation of endless traces that degrade performance.
Viewed through this lens, sleep begins to resemble something surprisingly familiar to computer scientists.
It looks like regularisation.
In machine learning, regularisation refers to a collection of techniques designed to prevent models from becoming excessively attached to specific training examples. Although the methods differ, their purpose is similar: maintain flexibility, preserve generalization and prevent runaway complexity. The goal is not to erase learning. The goal is to make learning sustainable.
The parallels with sleep are difficult to ignore.
During wakefulness, the brain accumulates modifications. Synaptic strengths change. Memories form. Predictions are updated. New experiences leave traces throughout multiple neural systems. Left unchecked, these changes could gradually increase the complexity and metabolic burden of the network. Sleep may function as a biological mechanism that prevents such accumulation from becoming pathological.
From this perspective, sleep is not the opposite of learning.
It is the second half of learning.
The distinction matters because it changes how we think about memory itself. Traditional accounts often portray memory as something stored, like files placed in an archive. Yet modern neuroscience increasingly suggests that memory is a dynamic process. Memories are not merely preserved; they are continuously reconstructed, integrated and reorganised. What matters is not simply retaining information but maintaining a system capable of using information effectively.
A useful analogy can be found in writing.
Imagine attempting to write a book without ever revising it. Every idea, every draft paragraph, every incomplete thought and every discarded sentence remains permanently embedded within the manuscript. At first, the accumulation seems productive because nothing is lost. Eventually, however, the text becomes unreadable. The problem is not insufficient content. The problem is the absence of selection.
Good writing depends as much on editing as on creation.
Perhaps learning follows a similar principle.
The sleeping brain may function partly as an editor.
It evaluates, reorganizes and selectively reduces the influence of accumulated traces, not because those traces were meaningless but because a system that retains everything with equal strength eventually loses the ability to distinguish what matters.
This perspective becomes even more interesting when viewed alongside predictive processing. In Friston's framework, the central task of the brain is to minimise prediction error while maintaining an accurate and flexible model of the world. Learning occurs when predictions are updated in response to evidence. Yet not every prediction error deserves equal influence. Some errors reveal important features of reality. Others are little more than noise.
A learning system must therefore solve two problems simultaneously.
First, it must detect prediction errors.
Second, it must determine how much weight to assign to those errors.
This second problem is often discussed in terms of precision. Precision reflects the estimated reliability or importance of a signal. A highly precise error should strongly influence future predictions. A low-precision error can safely be ignored.
The challenge is obvious. If a system assigns excessive precision to every error, it becomes unstable. Every anomaly demands a response. Every coincidence appears meaningful. Every deviation triggers learning. The result is a model that becomes increasingly fragmented and vulnerable to noise.
Seen from this perspective, the relationship between Friston's predictive framework and Tononi and Cirelli's synaptic homeostasis becomes striking. Although the theories emerged from different traditions and operate at different levels of explanation, they appear to address the same fundamental challenge.
Friston describes the problem in terms of prediction errors and precision weighting.
Tononi and Cirelli describe it in terms of synaptic strength and network stability.
In both cases, the underlying issue is remarkably similar.
How does a learning system prevent itself from becoming overwhelmed by its own adaptations?
How does it remain open to new information without treating every piece of information as equally important?
How does it preserve flexibility without sacrificing stability?
Sleep may be one of the answers.
The nightly reduction of overall synaptic strength proposed by SHY can be interpreted as a biological mechanism for controlling the accumulation of precision. Weak, transient and poorly supported signals lose influence. Strong and repeatedly reinforced patterns retain their relative advantage. The system returns to a state in which learning remains possible without becoming chaotic.
In this sense, sleep begins to look less like rest and more like calibration.
The brain spends the day building and modifying its model of reality. During the night, it may recalibrate the parameters that determine how strongly future experiences will shape that model. Without such recalibration, learning itself could become destabilising.
This idea also sheds light on a long-standing observation that has puzzled psychologists and neuroscientists alike. Sleep deprivation rarely produces a single dramatic cognitive deficit. Instead, it gradually impairs attention, emotional regulation, decision-making, memory and perception. The effects appear widespread because sleep may not be supporting one specific function. It may be maintaining the conditions that make many functions possible.
A brain deprived of sleep is not merely tired.
It may be operating with an increasingly dysregulated learning system.
Predictions become less reliable. Attention becomes less selective. Emotional reactions become more volatile. The boundary between signal and noise begins to blur. In extreme cases, prolonged sleep deprivation can even produce hallucinations and psychotic-like experiences, a fact that becomes particularly intriguing when viewed through the frameworks of predictive processing and synaptic homeostasis.
If sleep is involved in maintaining the stability of the predictive machinery itself, then disrupting sleep may gradually compromise the brain's ability to distinguish meaningful information from irrelevant fluctuations. The system remains active, but its calibration begins to drift.
The implications extend beyond basic neuroscience.
They suggest that any intervention designed to change the brain—whether psychotherapy, pharmacology, brain stimulation or behavioural training—may need to be understood as a two-stage process. The first stage occurs during wakefulness, when new experiences alter predictions, beliefs and neural connectivity. The second stage occurs during sleep, when those changes are evaluated, integrated and incorporated into the broader architecture of the mind.
In this view, sleep is not what happens after learning.
Sleep is part of learning.
It is not what happens after the brain changes.
It is part of the process by which change becomes stable.
And if that is true, then sleep is doing something even more important than consolidating memories. It is helping maintain the delicate balance that allows a predictive brain to remain both adaptable and coherent. It is preserving the conditions under which reality itself can continue to be modelled successfully.
Reality as a controlled hallucination
At first glance, the phrase controlled hallucination sounds self-contradictory. Hallucinations are usually associated with error, confusion or pathology, whereas reality is assumed to be stable, objective and shared. To place the two concepts side by side appears almost provocative.
Yet over the last several decades, neuroscience has gradually moved toward a view of perception that makes the phrase surprisingly difficult to dismiss.
The reason is simple. The brain never has direct access to the external world.
This statement is not philosophical. It is biological.
Everything the brain knows about reality arrives indirectly. Light reflected from objects reaches the retina as patterns of electromagnetic energy. Vibrations in the air become neural signals within the auditory system. Mechanical pressure activates receptors in the skin. Chemical molecules stimulate receptors in the nose and tongue. At no point does the brain encounter the world itself. It encounters only streams of information generated by sensory systems.
From the brain's perspective, reality is always inferred.
This inference problem is one of the most extraordinary challenges in biology. The sensory data reaching the nervous system are incomplete, noisy and often ambiguous. The same pattern of light can be produced by multiple objects. The same sound may have different causes. The same bodily sensation can reflect injury, illness, emotion or expectation.
Yet despite this uncertainty, our experience of reality usually feels effortless. We do not wake each morning confronted by an incomprehensible flood of sensory signals. We encounter a world that appears coherent, stable and meaningful.
The question is how.
The traditional answer suggested that perception works from the outside in. Sensory information arrives first, and the brain gradually builds an internal representation of reality from the incoming data.
Predictive processing turns this logic upside down.
According to predictive theories, perception begins not with sensation but with expectation. The brain continuously generates models of what it believes exists and then compares those models against incoming evidence. Reality is not assembled from scratch each moment. Instead, it emerges from a dialogue between prediction and correction.
This framework helps explain something that has long puzzled neuroscientists. The sensory information reaching the brain is often insufficient to determine a unique interpretation of the world. Yet perception rarely feels uncertain. The brain resolves ambiguity by relying on prior expectations.
When we recognise a face in poor lighting, understand speech in a crowded room or identify a familiar object from a partial view, we are witnessing prediction at work. The brain fills in missing information, generating a coherent interpretation from incomplete evidence.
In most cases, this process is extraordinarily successful.
Indeed, it may be one of the reasons humans can function so efficiently in a complex environment. Waiting for complete information before forming an interpretation would be computationally expensive and often dangerous. Prediction allows the nervous system to act quickly and effectively.
The cost of this efficiency, however, is that perception is never a simple copy of reality.
It is always an interpretation.
The philosopher Immanuel Kant argued centuries ago that human beings do not experience the world as it exists independently of observation. Rather, we experience the world as it appears through the structures of our own minds. Although modern neuroscience operates within a very different intellectual framework, predictive processing arrives at a surprisingly similar conclusion through entirely different means.
The world we experience is not reality itself.
It is reality as modelled by a predictive brain.
Dreams provide perhaps the clearest demonstration of this principle. During sleep, the external constraints on perception are dramatically reduced, yet conscious experience continues. The brain still constructs people, places, events and narratives. It still generates a world. The fact that this world can exist without corresponding external objects reveals something important. Conscious experience does not require direct access to reality. It requires a sufficiently coherent model.
Psychosis offers a different perspective on the same principle. In psychotic states, internally generated interpretations may begin to overpower external constraints. Predictions acquire excessive influence. Hallucinations and delusions can emerge when the balance between expectation and evidence becomes disrupted. The resulting experiences often feel compelling precisely because they arise from the same predictive machinery that ordinarily generates perception.
Neither dreams nor psychosis represent normal perception, yet both expose aspects of the process that are usually hidden.
They reveal that the brain is always constructing reality.
What changes is the degree to which that construction remains constrained by external evidence.
This brings us back to sleep.
The work of Tononi and Cirelli suggests that the brain cannot simply continue strengthening its internal models indefinitely. Learning itself creates a problem. Every experience modifies neural networks. Every successful prediction leaves traces within the system. Without mechanisms for recalibration, those traces would accumulate, increasing complexity, metabolic cost and vulnerability to noise.
Sleep may therefore perform a function far more fundamental than memory consolidation alone. It may help maintain the stability of the predictive machinery itself. By renormalising synaptic strengths, reducing the influence of weak traces and preserving the relative importance of stronger ones, sleep may prevent the brain from becoming trapped by its own history.
Seen in this light, sleep is not merely a biological necessity.
It is a safeguard against runaway model building.
The brain spends the day generating predictions, updating beliefs and adapting to experience. During the night, it may perform the equally important task of deciding which adaptations deserve to survive. The result is a system that remains capable of learning without becoming overwhelmed by learning.
What makes this synthesis so compelling is that it brings together several seemingly unrelated areas of research. Studies of dreaming, psychosis, perception, learning and sleep all begin pointing toward a common theme. The brain is not a passive observer of reality. It is an active constructor of reality. Conscious experience emerges from a constantly updated model that must remain both flexible and stable, capable of changing in response to new evidence while resisting disruption from noise.
Perhaps the most remarkable implication of this view is that reality and hallucination no longer appear as absolute opposites. Instead, they occupy different positions along a continuum of model generation and constraint.
At one end lies the dream, where internal models operate with minimal external correction.
At another lies psychosis, where the balance between prediction and evidence may become distorted.
At the centre lies ordinary waking perception, where internally generated models remain tightly anchored to sensory information.
The difference is not whether the brain constructs reality.
The difference is how effectively reality constrains the construction.
For centuries, people have asked what dreams mean. Modern neuroscience is beginning to ask a more unsettling question: what does waking reality mean?
Dreams demonstrate that the brain can generate entire worlds in the absence of the external environment. Predictive processing suggests that even during wakefulness, we never encounter reality directly, but rather experience a continuously updated model constrained by sensory evidence. Research on psychosis reveals what can happen when that constraint weakens. Research on sleep suggests that every night the brain performs maintenance on the machinery responsible for generating the model itself.
Taken together, these ideas lead to a conclusion that is both humbling and profound.
The world we experience is neither a direct copy of reality nor a pure invention of the mind. It is something in between: a living negotiation between expectation and evidence, prediction and correction, learning and stability. Reality is not passively received. It is actively constructed, continuously tested and endlessly revised.
Dreams matter because they expose this process. They allow us to glimpse the architecture that normally remains hidden behind the apparent solidity of everyday life. Every night, as the external world fades and internally generated realities emerge, the brain reveals something about its own nature.
It shows us that consciousness is not a mirror.
It is a model.
And reality, in the end, may be best understood not as a photograph of the world, but as the most successful prediction the brain has managed to maintain so far.
The future of psychiatry, therapy and consciousness
If the picture emerging from modern neuroscience is even approximately correct, its implications extend far beyond the science of sleep. What began as an attempt to understand dreams gradually leads to a radically different way of thinking about mental health, psychiatric treatment and even consciousness itself.
For much of the twentieth century, psychiatry was largely organized around symptoms. Depression was defined by low mood, loss of motivation and hopelessness. Anxiety disorders were defined by excessive fear and worry. Psychosis was defined by hallucinations, delusions and disturbances of thought. Treatments were evaluated primarily by their ability to reduce those symptoms.
This approach has undoubtedly saved lives and remains indispensable in clinical practice. Yet it often leaves a deeper question unanswered.
What exactly is being changed?
If the brain is fundamentally a predictive system—one that continuously constructs, updates and stabilizes a model of reality—then psychiatric disorders may not simply be collections of symptoms. They may represent disturbances in the processes through which reality itself is generated, interpreted and maintained.
From this perspective, depression is no longer merely sadness. Anxiety is no longer merely fear. Psychosis is no longer merely a false belief.
Each may reflect a particular way in which the brain's predictive machinery becomes biased, constrained or destabilised.
Consider depression. Traditional descriptions often emphasise mood, but depressive disorders involve much more than emotional suffering. They alter expectations about the future, interpretations of the present and memories of the past. A depressed person frequently experiences the world through a model that systematically predicts failure, loss, uncertainty or hopelessness. Positive information may be discounted, while negative information receives disproportionate weight.
Seen through the lens of predictive processing, depression begins to resemble a self-reinforcing model of reality. The problem is not simply that the individual feels bad. The problem is that the predictive system increasingly expects a world that justifies feeling bad.
Anxiety disorders may involve a related but distinct distortion. In anxious states, the brain appears to overestimate threat and uncertainty. Ambiguous situations are interpreted as potentially dangerous. Harmless bodily sensations become warning signals. Future events are modelled in increasingly catastrophic ways. The resulting fear feels rational because it emerges from a model that predicts danger everywhere.
What matters in both cases is that the person is not merely experiencing symptoms. They are inhabiting a reality generated by a particular predictive configuration.
This insight changes how we think about treatment.
Traditionally, therapy and medication are often discussed as separate interventions. Psychotherapy changes thoughts and behaviours. Medication changes neurochemistry. Yet from the perspective of predictive processing, both may ultimately be influencing the same system: the brain's internal model of reality.
Psychotherapy can be understood as a structured process for generating prediction errors. It exposes individuals to information that challenges existing models. It encourages alternative interpretations, new behaviours and experiences that contradict entrenched expectations. Effective therapy may not work because it provides insight alone. It may work because it creates conditions under which the brain can revise deeply held predictions about itself and the world.
Medication, meanwhile, may alter the flexibility of those models. Certain drugs appear to influence how strongly prediction errors are weighted, how rapidly learning occurs or how rigidly existing beliefs are maintained. The details differ across medications and disorders, but the general principle remains strikingly consistent.
Treatment may be less about removing symptoms and more about changing the conditions under which models can be updated.
This perspective becomes particularly interesting when we consider some of the most promising developments in modern psychiatry.
Over the past two decades, increasing attention has been directed toward interventions that appear to produce rapid changes in mental state. Ketamine, psychedelic-assisted therapies, transcranial magnetic stimulation (TMS) and several emerging treatments often seem capable of creating temporary windows of heightened plasticity. During these periods, established patterns of thought and perception may become more flexible than usual.
Researchers frequently describe these interventions as reopening critical periods of learning.
The language itself is revealing.
Rather than permanently fixing the brain, these treatments may temporarily increase its capacity for revision.
The implications are profound. If a therapy creates a window of plasticity, then the most important question may not be what happens during the intervention itself. The most important question may be what happens afterwards.
And this is where sleep re-enters the story.
If Tononi and Cirelli are correct, sleep is not merely a passive period that follows learning. It is an active process through which learning becomes integrated into the broader structure of the brain. New experiences, new beliefs and new emotional associations must ultimately be incorporated into a stable predictive model. That incorporation may depend heavily on what occurs during subsequent sleep.
This possibility suggests a very different way of evaluating psychiatric interventions.
Traditionally, clinicians ask whether a treatment improves symptoms during the day.
A predictive and sleep-based framework encourages an additional question.
What does the treatment allow the brain to do at night?
A therapy may generate important prediction errors. A medication may increase plasticity. A psychedelic experience may destabilise long-standing assumptions. Yet if the subsequent processes of consolidation, recalibration and network reorganisation are impaired, the long-term effects could differ dramatically from the immediate experience.
In this sense, change may be a two-stage process.
The first stage opens the system.
The second stage stabilises it.
One occurs primarily during wakefulness.
The other may occur largely during sleep.
This possibility remains an active area of research, and many questions remain unanswered. Nevertheless, it points toward a future in which sleep is no longer viewed as a secondary concern in psychiatry. Instead, it may become one of the central variables through which therapeutic change is understood.
The implications extend beyond clinical treatment.
They also reach into one of the oldest philosophical questions: what is consciousness?
For centuries, debates about consciousness were often framed in terms of subjective experience. Why does experience exist at all? How does awareness emerge from physical matter? Why does the brain produce a first-person perspective?
These questions remain difficult, but modern neuroscience has gradually shifted attention toward another mystery.
How does consciousness remain coherent?
At any given moment, the brain is integrating information from countless sources. Memories, expectations, sensory signals, emotions and bodily states are continuously interacting. Yet the resulting experience usually appears unified. We do not perceive isolated fragments of information. We perceive a world.
The predictive framework offers one possible explanation. Consciousness may not simply be awareness. It may be the ongoing process through which the brain generates and updates a model of reality. The self, the body and the external world become components of a single predictive structure designed to guide behaviour and reduce uncertainty.
Dreams reveal how that structure behaves when external constraints weaken.
Psychosis reveals how it behaves when predictive calibration becomes disrupted.
Sleep reveals how it may be maintained.
Taken together, these phenomena suggest that consciousness is not a static entity but a dynamic process. It is something the brain continuously does rather than something the brain simply has.
This perspective leads to a final and surprisingly humbling conclusion.
The greatest achievement of the human brain may not be intelligence, memory or reasoning.
It may be stability.
Every day, billions of neurons generate predictions, process errors, revise models and adapt to new information. Every night, those same networks appear to undergo processes that prevent adaptation from becoming chaos. Between these two phases emerges something so familiar that we rarely stop to notice it: a coherent reality.
Dreams, psychosis, learning and sleep all reveal different aspects of the same system. They remind us that reality is not something passively received from the outside world. It is actively constructed by a biological organ that must constantly balance flexibility against stability, learning against coherence and imagination against evidence.
The future of psychiatry may ultimately depend on understanding this balance more deeply. The future of consciousness research may depend on it as well.
And perhaps the future of neuroscience itself will continue to circle back to the same surprising place where this story began: a sleeping brain, alone in the dark, constructing a world from within.
Comments
Post a Comment