The Predictive Brain: Predictive Processing and the Illusion of Perception

The Hook: A Mask That Refuses to Be Hollow

Take an ordinary face mask — the kind worn at Halloween — and turn it around so you are looking into its hollow inner side. The nose points away from you, the eye sockets curve inward. You know with absolute certainty that you are looking at the concave back. And yet: the moment there is enough light and a little distance, the image flips. Your brain refuses to see the hollow mask as hollow. It shows you a normal, outward-bulging face that even seems to track your movements. Healthy observers miscategorize the hollow mask as convex roughly 99 percent of the time — even when they know better.

This is the famous hollow-mask illusion, and it is more than a fairground trick. It is a window into one of the most influential ideas in modern cognitive science: your perception is not a faithful recording of the world flowing into you from outside. It is a prediction your brain generates from the inside out — a best guess about the causes of your sensory input, constantly compared against the data actually arriving and corrected accordingly. In the hollow mask, the brain's lifelong experience with faces ("faces bulge outward") is so overwhelmingly strong that it simply overrides the unambiguous depth cue on the retina.

This view is called predictive processing (often also predictive coding). Its core can be summarized in a single, initially disconcerting sentence: the brain is not a passive receiving station for sensory data but a prediction machine whose supreme goal is to keep the difference between what it expects and what it actually receives as small as possible. What you call "perception," on this reading, is the prediction that currently fits best — the neuroscientist Anil Seth aptly calls it a "controlled hallucination."

Why should this interest you — beyond the intellectual appeal? Because over the past two decades this framework has risen to become the most ambitious candidate for a unified theory of the brain. It promises to explain perception, action, attention, learning, and even psychiatric conditions such as autism, schizophrenia, and the effects of psychedelics under a single mathematical principle. This article takes you the whole way: from the philosophical roots in Helmholtz, through the neural mechanics of prediction errors and the decisive role of "precision," to the clinical applications — and finally to the honest question of how well established all of this really is.

Part 1: The Inversion of Perception

The Intuitive View — and Why It Misleads

The naive picture of perception is a one-way street from outside to inside. Light strikes the retina, is translated into nerve impulses, travels up the visual pathway, is assembled step by step into ever more complex features — edges, shapes, objects, meanings — until, somewhere at the very top, a finished picture of the world emerges. On this view the brain is essentially reactive: it receives the world.

Predictive processing reverses this flow. The main direction of information processing, the thesis goes, runs not from the bottom up but from the top down. Higher brain areas continuously send predictions to the lower-lying ones: "This is what the signal that is about to arrive should look like." The ascending pathways then carry not the raw sensory signal but only its surprising remainder — the difference between prediction and reality. This difference is called prediction error. What travels upward, then, is not "what I see" but "where I was wrong."

The payoff of this architecture is, first, an information-theoretic one. If a system predicts the future well, it need only communicate the deviations, not the entire, highly redundant data stream. This is exactly the principle engineers know from data compression: a video codec does not retransmit every full frame either, but only the changes relative to the predicted frame. The brain, the idea goes, learned the same lesson millions of years ago.

Helmholtz and "Unconscious Inference"

The idea is surprisingly old. As early as 1867, the polymath Hermann von Helmholtz coined the term unconscious inference. Helmholtz recognized a fundamental problem: the image on the retina is ambiguous. One and the same flat pattern can be produced by infinitely many different object configurations in the three-dimensional world. The brain must infer, from an ambiguous effect, its most probable cause — and this inference, according to Helmholtz, happens automatically, pre-rationally, and reflexively, drawing on a store of experience about how the world typically is.

This formulation is essentially already the modern picture: the perceptual system is a statistical inference engine that infers the hidden causes from ambiguous data. It is no coincidence that the pioneering computational-model work by Peter Dayan, Geoffrey Hinton, and colleagues from the 1990s bears the title The Helmholtz Machine. The psychologist Richard Gregory revived the idea in the 1970s and demonstrated it strikingly with that very hollow mask: perception as hypothesis formation, as "perceiving is guessing."

Rao and Ballard: The Model Becomes Concrete

It was a long way from the philosophical thought to a precise neural model. The decisive step was taken in 1999 by Rajesh Rao and Dana Ballard in a paper in Nature Neuroscience that is cited thousands of times to this day. They described a hierarchical network in which feedback connections from a higher to a lower visual area carry predictions about the neural activity there, while the feedforward connections relay only the residual error between prediction and actual activity.

The remarkable part: when Rao and Ballard trained this network on natural images, the model neurons spontaneously developed sensitivity profiles that strikingly resembled the simple cells in the visual cortex. And a subset — those neurons that carry the prediction error — showed so-called extra-classical receptive-field effects such as end-stopping: phenomena previously regarded as purely feedforward properties. Rao and Ballard showed that these effects arise naturally in a system that tries to predict images efficiently and hierarchically. For the first time, the abstract idea had a plausible neural mechanism.

Friston and the Free Energy Principle

The most sweeping — and most contested — generalization comes from the British neuroscientist Karl Friston. His free energy principle claims nothing less than that every self-maintaining system — every cell, every brain, every organism — fundamentally minimizes a single quantity: variational free energy, an information-theoretic upper bound on "surprise" (formally: the negative log probability of the sensory data under the system's internal model).

Sounds abstract, but it has a vivid core. A living being exists only as long as it remains within a narrowly bounded set of states — a fish in water, not in air; a body temperature around 37 degrees, not 50. From the organism's point of view, these are the "expected," unsurprising states. To minimize surprise, then, ultimately means to maintain one's own existence. Predictive processing, in Friston's picture, is the concrete neural implementation of this general imperative: the brain minimizes prediction error because that is the practically computable version of "minimize surprise." An important distinction: the free energy principle is the overarching theoretical framework, while predictive coding is a concrete theory of how neurons might implement it. We will return to the legitimate criticism of the strong version.

Part 2: How the Machine Works

The Hierarchy of Predictions

Picture the cortex as a tower of levels. Each level tries to predict the activity of the level beneath it. The lowest level is the sensory input itself. Between any two neighboring levels, two streams flow in opposite directions:

The descending stream carries predictions. Higher levels encode more general, more abstract, temporally more stable hypotheses ("there is a cat in front of me"); lower levels encode more concrete, faster-changing details (edges, contrasts, motion). The higher level tells the lower which pattern it should expect.

The ascending stream carries prediction errors. If the prediction is correct, the error is small, and almost nothing travels upward — the higher level is confirmed in its hypothesis and stays quiet. If the prediction is wrong, an error signal arises that is carried upward and forces the higher level to revise its hypothesis.

Perception is thus a continuous negotiation between top and bottom, until the system settles on the hypothesis that minimizes the total error across all levels. This winning hypothesis is your conscious percept. That is why one can say: you do not see the world but the model of the world that currently best explains your sensory data.

The Decisive Dial: Precision

Pure "error minimization" would be dangerously naive. Not every sensory stimulus is equally reliable. In thick fog the visual input is noisy and should be given less weight; a bang in the still of night is a highly reliable signal. So the brain must not only compute errors but also estimate how seriously to take each error.

This weighting is called precision — in the statistical sense the reciprocal of the expected variance, that is, a measure of reliability. Precision is arguably the most important dial in the whole system. A prediction error with high precision is heard loudly and forces a revision of the hypothesis; an error with low precision is dismissed as noise and ignored. On this reading, attention is nothing other than the targeted increase of the precision of particular prediction errors. When you concentrate on a faint tone, you turn up its precision dial — its errors gain more influence over what you perceive and do.

At the neural level it is thought that neurotransmitters such as dopamine, acetylcholine, and noradrenaline modulate this precision weighting — the synaptic "gain" of the error-carrying neurons. This connection between an abstract computational quantity and concrete neurochemistry is what makes the framework clinically tractable, as we shall see shortly.

Four Pieces of Evidence from the Lab

How does one become convinced that the brain actually works this way? Four lines of observation are repeatedly cited — though they should be read with due caution.

First, repetition suppression: when a stimulus is repeated, the neural response to it declines. From the predictive-processing point of view this is natural — the repeated stimulus is increasingly well predicted, the prediction error shrinks, so the ascending signal falls silent. Importantly, this suppression is context-dependent: when the repetition itself is expected, it is stronger than when it comes as a surprise — a detail that a pure fatigue account struggles to explain.

Second, the mismatch negativity (MMN): present a series of identical tones and then a deviant one, and the outlier produces a characteristic negative deflection in the EEG. The unexpected stimulus triggers a large prediction error — exactly what the theory predicts. MMN is regarded as one of the most robust electrophysiological correlates of prediction errors and is also found in the visual system.

Third, binocular rivalry: show each eye a different image and you do not see both simultaneously; instead the percept flips back and forth between them every few seconds. Predictive processing explains this as a contest between two hypotheses, neither of which can permanently explain the input.

Fourth, the top-down effects themselves: in the hollow mask, imaging studies show strengthened descending coupling from higher areas (intraparietal sulcus) to lower ones (lateral occipital cortex) — precisely the direction in which the theory locates the predictions.

An honest caveat belongs here: much of this evidence is indirect. A core assumption — that expected stimuli are encoded more efficiently in the brain than unexpected ones — is astonishingly hard to test directly, and some studies find that repetition and change can produce nearly identical signals. Predictive processing is a very well-fitting but not yet conclusively proven framework.

Part 3: Perception Is Only Half the Story — Active Inference

So far we have considered only one way to reduce prediction error: change the hypothesis until it fits the input. That is perception. But there is a second, equally effective way: change the input until it fits the hypothesis. That is action.

This symmetry is perhaps the most elegant idea in the whole framework and is called active inference. Suppose your brain generates the prediction "my hand is about to touch the coffee cup." At first this is a prediction error — the hand is not there yet. You can resolve this error in two ways: either you revise the prediction (the hand stays put), or you move the hand so that sensory reality catches up with the prediction. Action thus becomes a self-fulfilling prophecy: the brain predicts the sensory consequences of a movement and then lets the reflex arcs of the spinal cord turn this prediction into reality by dissolving the error between expected and actual body posture.

Perception and action, on this view, are not separate modules but two complementary strategies for the same task: to reduce prediction error. One fits the model to the world, the other fits the world to the model. And this is exactly where precision plays the key role again: for a movement to occur at all, the brain must temporarily turn down the precision of the sensory errors coming from the not-yet-moved hand — it must ignore the counter-evidence "but my hand is still lying still," or the prophecy would never fulfill itself. This "sensory attenuation" incidentally explains why you cannot tickle yourself: the self-generated stimuli are precisely predicted and are dampened.

Part 4: When Prediction Goes Off the Rails — The Clinic

Perhaps the strongest reason to take the framework seriously is its tractability for psychiatry. Many disorders can be elegantly reformulated as disorders of precision — as a miscalibrated weighting between predictions (priors) and prediction errors.

Schizophrenia and Psychosis

Why should a person hear a voice that is not there? A predictive-processing reading runs: people with pronounced hallucinations overestimate the precision of their predictions at higher levels. When a strong expectation ("someone is speaking there") is weighted with very high precision, it can override the missing sensory input — the brain "completes" a perception for which it has no evidence at all. Fittingly, studies show that hallucinators report false tones more often than non-hallucinators in ambiguous listening situations. Delusions, in turn, can be read as inappropriately rigid prior beliefs that are no longer corrected by normal counter-evidence. Since dopamine is regarded as a modulator of precision, this picture even builds a bridge to the well-known dopamine hypothesis of psychosis.

Autism: The Aberrant Precision Account

For the autism spectrum, researchers such as Rebecca Lawson, Karl Friston, and colleagues have proposed a reading under the heading of aberrant precision. Simplified: if sensory prediction errors are chronically assigned too high a precision, the world is experienced as overwhelmingly detailed and unpredictable. Every small deviation screams for attention instead of being smoothed away as negligible noise. This could straightforwardly explain why many autistic people show an exceptional perception of detail but also a pronounced sensory overload and a strong need for predictable, unchanging environments. I am of the opinion that this reading explains a great deal, but one should treat it with caution: autism is heterogeneous, and a single computational principle will hardly do full justice to the diversity of experiences.

Psychedelics: The Relaxation of Beliefs

One of the most fascinating applications concerns the effects of classic psychedelics such as psilocybin and LSD. The REBUS model (Relaxed Beliefs Under pSychedelics) by Robin Carhart-Harris and Karl Friston proposes that these substances — mediated via the serotonin 2A receptor — relax the precision of high-level prior beliefs. Figuratively speaking, the "energy landscape" of beliefs becomes flatter: deeply entrenched, rigid expectations lose their grip, and the prediction error rising from below can once again flow upward more freely and reshape the beliefs.

This provides a mechanistic language for phenomena that are otherwise hard to grasp: the dissolution of the sense of self, the reappraisal of stuck patterns of thought, the potentially therapeutic effect in depression. If depressive rumination loops are understood as excessively precise, dysfunctional prior beliefs, then their temporary relaxation — followed by a phase of increased plasticity — would be a plausible mechanism of action. It should be emphasized that REBUS is a theoretical model that is actively being tested and also criticized; more recent work (such as the ALBUS model from 2025) proposes refinements. But the very fact that perception, psychosis, autism, and psychedelic effects can be discussed in the same conceptual language — prediction, error, precision — is itself a strong argument for the reach of the framework.

Part 5: A Compass Through the Concepts

For orientation, a compact overview of the core concepts:

The common thread: there is only one currency — the precision-weighted prediction error — and only one task — to minimize it. Everything else, from perception through attention to movement, are different ways of carrying out this one computation.

Part 6: The Honest Ledger — How Well Established Is All This?

A framework that explains everything must face the critical questions. Three are especially important.

First, the charge of unfalsifiability, aimed above all at the strong version, the free energy principle. If literally every behavior — perception, action, learning, even cell death — can be rewritten as "minimization of free energy," what would the theory actually forbid? A theory compatible with every conceivable finding, strictly speaking, predicts nothing. Defenders reply that the principle is not an empirical hypothesis at all but a mathematical framework — comparable to the principle of least action in physics. Critics from the ecological-enactive camp object that the anticipating brain is "not a scientist": it does not necessarily build an explicit model of the world but is simply adaptively coupled with it.

Second, the already-mentioned indirectness of the evidence. Repetition suppression and mismatch negativity are compatible with predictive processing, but in part they can also be explained by simpler mechanisms such as neural adaptation. Direct proof that the cortex maintains separate populations for predictions and for errors is difficult and remains the subject of active research.

Third, the question of the diversity of implementations. "Predictive coding" now designates a whole family of models that differ in important details. Some findings support one variant and contradict another. This is not a flaw — that is how healthy science works — but it counsels caution toward the narrative of the one grand theory of the brain.

Interestingly, a circle closes here toward artificial intelligence. Predictive-coding networks can be formulated as a biologically plausible alternative to classical backpropagation learning, and the "world models" of modern AI research — systems that predict the next observation in order to act — are immediate relatives of active inference. Whether or not the brain computes in exactly this way: the idea that intelligence is at its core prediction today shapes both disciplines in equal measure.

The Central Takeaway

The central insight of predictive processing is at once disconcerting and practical. Disconcerting: your experienced reality is not the world but the model of the world that currently best predicts your sensory data — a controlled hallucination, constantly compared against reality but never identical with it. What you take for immediate perception is always already interpretation.

It becomes practical through a single dial: precision, that is, the question of how much weight you give your expectations relative to the incoming evidence. Whoever sets their priors too precisely goes blind to contradiction — they see only what they expect, whether in the hollow mask, the entrenched prejudice, or the depressive rumination loop. Whoever sets them too loosely drowns in noise and finds no foothold. Wise thinking, in this picture, is above all good precision management: the willingness to give your prediction errors a hearing when they are reliable — especially when they contradict your own expectation. That, soberly regarded, is a neuroscientific reformulation of a very old virtue: attention to what surprises us.

A Question to Reflect On

If your perception is always a blend of what arrives and what you expect — how much of what you took to be "obvious" today was a precise observation, and how much was a particularly precisely weighted expectation that never let the contradiction through in the first place?

Cross-References in the Vault

• Desirable Difficulties: Why Effortless Learning Deceives and Memory Lives on Effort – Learning as adjustment of the generative model; the prediction error is the engine of relearning.
• Three Pages Against 2,000 Years: The Gettier Problem and What Knowledge Really Is – If perception is inference, the epistemological question of justified, true belief becomes even thornier.
• The Two Boxes: Newcomb's Paradox and the Battle Over Rational Choice – Bayesian priors and expected values as a shared language of perception and decision.
• The Ghost in the Machine: How to Read a Neural Network From the Inside – Prediction machines made of silicon: what artificial and biological neural networks share about representation.

Sources

• Rao, R. P. N. & Ballard, D. H. (1999): Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects. Nature Neuroscience 2, 79–87. https://www.nature.com/articles/nn0199_79
• Friston, K. (2010): The free-energy principle: a unified brain theory? Nature Reviews Neuroscience 11, 127–138. https://www.nature.com/articles/nrn2787
• Millidge, B., Seth, A. & Buckley, C. L. (2021): Predictive Coding: a Theoretical and Experimental Review. arXiv:2107.12979. https://arxiv.org/pdf/2107.12979
• Stefanics, G., Kremláček, J. & Czigler, I. (2014): Visual mismatch negativity: a predictive coding view. Frontiers in Human Neuroscience 8:666. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4165279/
• Lawson, R. P., Rees, G. & Friston, K. J. (2014): An aberrant precision account of autism. Frontiers in Human Neuroscience 8:302. https://pmc.ncbi.nlm.nih.gov/articles/PMC4030191/
• Smith, R., Badcock, P. & Friston, K. J. (2021): Recent advances in the application of predictive coding and active inference models within clinical neuroscience. Psychiatry and Clinical Neurosciences 75, 3–13. https://onlinelibrary.wiley.com/doi/10.1111/pcn.13138
• Carhart-Harris, R. L. & Friston, K. J. (2019): REBUS and the Anarchic Brain: Toward a Unified Model of the Brain Action of Psychedelics. Pharmacological Reviews 71(3), 316–344. https://pmc.ncbi.nlm.nih.gov/articles/PMC6588209/
• Bruineberg, J., Kiverstein, J. & Rietveld, E. (2018): The anticipating brain is not a scientist: the free-energy principle from an ecological-enactive perspective. Synthese 195, 2417–2444. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6438652/

← All articles