Predictive Processing in the Brain: Models, Evidence, and Debate

4 days ago

21 min read

Author: Ukrainian Psychological Hub · Published: October 12, 2026 · Editorial Policy

Predictive processing describes a family of theories in which nervous systems use learned regularities and current context to anticipate sensory information, interpret ambiguous inputs, and adjust their estimates when outcomes differ from expectations. It is one of contemporary neuroscience’s most productive proposals for connecting perception, learning, and action. Its central scientific question is how predictions are represented, compared with signals, and updated in actual neural circuits.

The evidence establishes that expectations influence perception and many neural responses. It does not yet establish one universal predictive-coding algorithm operating throughout the brain. Two major syntheses reach this nuanced conclusion: Hodson, Mehta, and Smith (2024) identify modest empirical support for specific predictive-coding mechanisms and alternative explanations for some positive findings; Furutachi and Hofer (2026) emphasize the need to determine what experimentally observed sensory-error-like signals actually encode. This distinction organizes the entire field.

What Is Predictive Processing in Neuroscience?

Predictive processing is a broad theoretical framework that treats perception and other neural operations as partly dependent on predictions about sensory signals or their underlying causes. These predictions may reflect the statistical structure of the environment, previous learning, the context of a task, or the anticipated sensory consequences of action. Incoming evidence can confirm, constrain, or revise the system’s working estimates.

A prediction in this context is a modeled expectation about a state or signal. It need not be conscious, verbal, or explicitly directed toward a future event. When a listener hears an incomplete phrase, contextual knowledge can alter the probabilities assigned to different sounds or words. When someone moves their eyes, motor and sensory systems can anticipate the consequences of that movement. Both examples involve predictive information, but they need not depend on the same algorithm.

The term encompasses multiple computational levels. At the functional level, a theory may claim that organisms benefit from exploiting statistical regularities. At the algorithmic level, a model specifies representations and update rules. At the biological level, a mechanistic theory identifies the neurons, synapses, connections, temporal dynamics, and causal interventions that implement those rules. Showing that a model predicts behavior addresses a different level of evidence from demonstrating its cellular implementation.

This framework belongs within the broader science of neural circuits and distributed brain networks. It does not replace established research on sensory physiology, attention, learning, or behavior. Its value lies in generating additional testable explanations for those observations.

Why Would the Brain Make Predictions?

Sensory signals are incomplete and dependent on context. Light reaching the retina varies with illumination, viewing angle, movement, and occlusion. Sound is shaped by background noise, distance, and preceding speech. An organism that draws on regularities learned from experience can often interpret these signals more efficiently than one that treats every moment as isolated.

Prediction can also help coordinate action. Movement changes the incoming sensory stream; without anticipating some of those changes, a system would face a moving target. Sensory predictions can support perceptual stability, motor control, adaptation, and the detection of informative departures from the usual pattern.

These are functional reasons that prediction could be useful. They do not prove that particular cells carry a numerical subtraction between predicted and actual input. As Teufel and Fletcher (2020) explain, prediction takes multiple forms in the nervous system. Learned local dynamics, context-dependent processing, efference copies, and hierarchical feedback can all generate expectation-sensitive neural activity.

The Historical Development of Predictive Models

The idea that perception uses prior knowledge has a long history in theories of perceptual inference. Twentieth-century research on sensory coding, statistical learning, feedback connections, and efficient representations supplied more explicit neural questions. In 1999, Rao and Ballard presented an influential model of vision in which higher visual levels predicted activity at lower levels, while feedforward connections transmitted residual errors. Their simulations explained selected contextual response properties, demonstrating computational plausibility rather than proving that every cortical circuit uses that architecture.

Friston’s 2005 account connected cortical responses with hierarchical inference, and his later free-energy formulation (2010) advanced a wider family of arguments about perception, learning, and action. Bastos and colleagues (2012) mapped predictive-coding ideas onto proposed cortical microcircuits, drawing attention to feedforward and feedback anatomy and physiological rhythms. Keller and Mrsic-Flogel (2018) articulated specific experimental tests of cortical predictive-processing hypotheses.

These traditions differ in what they specify. A predictive framework can be useful as a formal model while its proposed neural implementation remains unsettled. Indeed, Spratling’s algorithmic review (2017) identifies several distinct algorithms described under the same name.

Predictive Processing, Predictive Coding, Bayesian Inference, and Active Inference

Predictive processing: the broad research framework

Predictive processing names the family of approaches emphasizing expectations, uncertainty, internal models, and their interaction with sensory information. Depending on the account, it may concern visual and auditory perception, action, interoception, learning, or decision-making. Its breadth explains its influence and creates a challenge: different experiments can test different propositions while all being reported as evidence for the same broad framework.

Predictive coding: an algorithmic proposal

Predictive coding is more specific. In a conventional hierarchical version, higher levels send predictions toward lower levels, and lower levels transmit quantities related to discrepancies between predicted and observed signals. Local updating reduces a selected measure of mismatch. Particular models additionally assign identifiable populations to prediction and error signaling and propose specific circuit dynamics.

A mismatch response is not automatically evidence for this complete architecture. The circuit must have the relevant computational organization, and that organization must outperform plausible alternatives in tests designed to distinguish them.

Bayesian inference: a computational description of uncertainty

Bayesian inference describes how estimates can depend on prior probabilities and the likelihood of observations. Conceptually, posterior belief about a cause is proportional to prior belief about that cause multiplied by the likelihood of the sensory evidence given that cause. Precision, often connected to inverse variance in Gaussian models, expresses how much uncertainty a source of information carries.

Bayesian behavior can arise from mechanisms other than predictive coding. Predictive coding can also serve computations that are not fully Bayesian. This separation is a central argument of Aitchison and Lengyel (2017). A behavioral match to Bayesian integration shows something about the functional pattern of decisions; it does not identify the neuronal message-passing algorithm.

Active inference: models that also select actions

Active-inference accounts extend inferential modeling into action and policy selection. Depending on the formulation, an agent may update beliefs, act to sample informative evidence, or change its sensory exchanges with the environment in ways that fulfill model-based expectations. This differs from a narrow perceptual predictive-coding hypothesis.

Some active-inference models fit behavioral data well, but broad empirical confirmation remains open. Hodson and colleagues (2024) evaluate predictive coding and active inference separately precisely because successful model-fitting in one domain cannot validate another model’s neural mechanism.

The free-energy principle: an ambitious formal proposal

The free-energy principle proposes mathematical relationships among model evidence, uncertainty, inference, and adaptive organization. In variational treatments, free energy provides a tractable bound related to surprisal under specified assumptions. This use of free energy should not be conflated with everyday surprise, metabolic energy, or a directly measured neurotransmitter concentration.

The scientific interpretation depends on the exact assumptions, generative models, and measurable predictions being tested. A broad variational formulation does not automatically identify how human perception works or establish a unique biological circuit. See the original Friston review (2010) and the distinction between computation and implementation emphasized by Aitchison and Lengyel (2017).

How Does a Hierarchical Predictive-Coding Model Work?

Imagine an observer trying to recognize an object in partial shadow. Visual input contains edges and contrast patterns. A model estimates possible sources of those patterns, generates expectations about the signals each source would produce, and revises estimates when the actual observations conflict with them.

In a simplified hierarchy, one level represents relatively local sensory features, while another represents broader patterns or probable causes. Predictions descend from a higher level; information about discrepancies ascends. Iterative exchange can improve agreement between the model and sensory data. The hierarchy is a computational organization, not a claim that one anatomical location knows the truth while another merely reports error.

An elementary prediction-error term can be written as observed signal minus predicted signal. More realistic schemes use nonlinear generative models, probabilistic states, gain control, and separate timescales for inferring current states and learning model parameters. This distinction matters because multiple algorithms can produce seemingly similar neural activity.

An error signal can be positive or negative, feature-specific, or expressed in neuronal population patterns. It need not appear as a single burst of neural firing, and a large response to an unexpected stimulus does not uniquely reveal an encoded subtraction.

Neural systems also operate recurrently. Signals travel through cortical and subcortical loops; local circuits contain excitatory and inhibitory interactions; sensory information changes during action. Realistic predictive models must explain timing, connectivity, population-level representation, and behavior together. Rao and Ballard (1999) offered a foundational computational example, while Bastos et al. (2012) and Keller and Mrsic-Flogel (2018) examine proposed biological implementations.

Prediction Error: What Does It Actually Mean?

Prediction error means a discrepancy between some expected quantity and an observed or inferred quantity under a defined model. Its meaning depends on what was predicted. Sensory prediction errors concern stimuli or representations; reward prediction errors concern outcomes relative to expected reward; motor-related mismatches concern predicted consequences of movement. Treating these as identical because they share a phrase erases crucial computational differences.

The reward prediction error literature examines particular learning signals and their relationships with dopamine and behavior. It is a neighboring EN-30 topic, not a demonstration that all sensory predictive coding depends on dopamine. For the wider circuitry of reinforcement and motivation, see the brain reward system.

A useful experimental question asks exactly which variable a neural response predicts: a stimulus identity, a probability, a sensory feature, movement-dependent feedback, a task outcome, or a change in confidence. Without a well-defined variable and task, the label prediction error may merely redescribe the observation.

Precision Weighting: Why Not All Errors Matter Equally

Many probabilistic predictive-processing models assign different weights to evidence depending on uncertainty. A reliable signal can exert a greater influence on updated estimates than a noisy signal. The relative weighting of prior expectations and current observations depends on the assumptions of the model and the structure of the task.

In Gaussian versions, precision is mathematically related to inverse variance. Neuroscientific proposals often connect changes in gain or attention with precision weighting. But neither attention nor a measured increase in neural activity is itself a direct readout of numerical precision. Establishing that link requires an explicit model and a way to test its predictions against other accounts.

Consider hearing a familiar word in heavy background noise. Linguistic context might become more influential as acoustic information becomes less reliable. In clean speech, the sensory evidence may constrain interpretation more sharply. Whether a particular participant performs Bayesian precision weighting, uses learned associations, or combines several strategies must be decided experimentally. The review by de Lange, Heilbron, and Kok (2018) evaluates the many ways expectations can shape perception.

What Brain Structures and Networks Are Involved?

There is no single predictive-processing center in the brain. Visual and auditory systems, thalamocortical pathways, sensorimotor loops, association cortex, and subcortical circuits can all participate in different prediction-sensitive computations. Models usually invoke distributed, recurrent interactions spanning different spatial and temporal scales.

In vision, research often examines primary visual cortex and higher visual areas, asking whether feedback conveys contextual expectations and whether activity under expected stimuli differs from responses to violations. In audition, studies examine cortical responses to tone regularities, sequence structure, and omissions. Sensorimotor research examines how self-generated movements change sensory input.

Predictive theories make anatomical predictions when they specify cellular populations, laminar connectivity, or directions of information flow. Proposed canonical microcircuit accounts assign computational roles to certain feedforward and feedback pathways. Those assignments are valuable hypotheses, yet circuit organization varies across species, cortical areas, tasks, and development. A single simplified laminar diagram should not be treated as a universal anatomical fact.

This is why the neural circuits overview and the general brain-behavior explanation provide a more stable biological foundation than assigning complex psychological functions to isolated regions.

Experimental Evidence: What Has Actually Been Found?

Visual expectations and representational sharpening

A controlled human fMRI study by Kok, Jehee, and de Lange (2012) found that valid expectations reduced the response amplitude in primary visual cortex while improving stimulus-related representation measured using multivariate analyses. The result shows that lower average activity and more informative patterns can coexist. It supports expectation-sensitive visual processing and is consistent with some predictive models.

The inference has limits. BOLD signals reflect hemodynamic changes linked to neural activity, not direct measurement of a prediction-error neuron. Improved decoding does not establish that the represented information traveled downward as a prediction or upward as an error. A mechanistic explanation requires stronger evidence about direction, circuitry, and computation.

Auditory mismatch negativity and violations of regularity

The mismatch negativity, often measured in EEG or MEG studies, is an electrophysiological response associated with deviations from auditory regularities. It has been studied for decades and is a major source of evidence that the brain registers changes in the structure of sound sequences. The Garrido et al. review (2009) explains why adaptation-based and model-adjustment explanations must both be considered.

A classic complication is that a rare sound differs not only in predictability but also in how frequently its responsive neurons were stimulated. Neural adaptation or stimulus-specific refractoriness can therefore generate an apparent surprise effect even without the hierarchical error representation proposed by a predictive-coding theory.

Wacongne and colleagues (2011) used a hierarchical novelty paradigm with EEG and MEG to separate aspects of repetition and violated expectations. Their findings support sensitivity to structured regularities at multiple levels. They do not by themselves identify a uniquely specified message-passing algorithm for the entire cortex.

Sensorimotor mismatch in animal experiments

Keller, Bonhoeffer, and Hübener (2012) reported sensorimotor mismatch responses in the primary visual cortex of behaving mice. This work is important because it relates neural responses to the expected visual consequences of movement, creating opportunities to manipulate motor-sensory relationships under controlled conditions.

The study concerns mice, an experimentally defined visuomotor task, and particular neuronal responses. It cannot establish by itself that human perception, self-awareness, or psychiatric symptoms use precisely the same circuit. Animal models are especially informative when the causal manipulations and relevant behavioral variables are explicitly identified.

Cortical microcircuits and neural implementation

Specific implementations propose patterns of feedforward and feedback signaling, interneuron contributions, gain modulation, and population dynamics. Bastos et al. (2012) framed a testable cortical microcircuit account; Keller and Mrsic-Flogel (2018) discuss how circuitry might confirm or falsify proposed architectures.

Evidence for recurrent signaling, context sensitivity, or laminar asymmetries supports constraints shared by many models. To validate a specific predictive-coding architecture, researchers must demonstrate the proposed representational content and causal information flow rather than infer them from anatomical resemblance alone.

What the major evidence reviews conclude

Walsh et al. (2020) evaluate the neurophysiological predictions and methodological difficulties associated with predictive-processing accounts of perception. Hodson et al. (2024) conclude that empirical support for predictive coding is modest and that alternatives can explain some supposedly distinctive results. The 2026 Annual Review of Neuroscience assessment revisits decades of sensory prediction-error research and foregrounds inconsistency in definitions and interpretations.

The cumulative finding is scientifically meaningful: expectation-dependent behavior and neural activity are well supported, and several model-specific predictions receive experimental support. The further claim that a single canonical predictive-coding algorithm explains all cortical function remains an active research hypothesis.

How Researchers Test Predictive-Processing Models

An experimental test becomes persuasive when it separates competing explanations rather than merely producing a pattern the theory can accommodate.

Behavioral experiments and psychophysics

Researchers manipulate probabilities, contextual cues, stimulus reliability, and feedback to measure discrimination, reaction times, perceptual reports, or confidence. These experiments establish how behavior depends on expectations. They can also compare quantitatively specified models and evaluate generalization to new conditions.

Behavioral responses rarely identify the neural algorithm by themselves. An observer could exhibit similar choices through explicit reasoning, statistical learning, heuristic strategies, or the proposed inferential mechanism. Predefined alternative models and out-of-sample prediction improve interpretability.

EEG and MEG

EEG records electrical voltage differences at the scalp generated largely by synchronized postsynaptic neuronal currents. MEG records weak magnetic fields associated with coordinated neural currents. Both offer temporal resolution suitable for studying when neural responses to expected and unexpected signals diverge.

Neither method directly measures the content of a neural prediction or reveals a unique generator from sensor data alone. Source localization is an inverse problem constrained by assumptions, and components such as mismatch negativity require task-specific interpretation.

Functional MRI

fMRI commonly examines the blood-oxygen-level-dependent signal, an indirect hemodynamic correlate of neural activity. Experimental contrasts can identify regional responses associated with expectation, surprise, or probabilistic context. Multivariate analyses can test whether activation patterns carry information that distinguishes conditions.

An fMRI activation decrease is not a direct measurement of reduced prediction error; an activation increase is not proof of encoded surprise. Neurovascular coupling, task design, analysis choices, and multiple possible neural processes matter. A concise comparison of what each technique measures appears in Brain Imaging Methods: What MRI, fMRI, PET, and EEG Measure.

Intracranial recordings and causal interventions

Animal electrophysiology, intracranial recordings in eligible clinical contexts, and carefully controlled stimulation can provide additional spatial, temporal, and causal specificity. Perturbing a predicted source of feedback or a candidate error population while measuring subsequent behavior can distinguish mechanisms more directly than correlational observation alone.

Even a causal effect of a circuit perturbation does not automatically establish that the affected circuit performed predictive coding. The manipulation must discriminate the proposed computation from alternative causal functions of the same population.

Computational model comparison

Researchers fit models of different complexity, simulate neural and behavioral trajectories, and evaluate predictions on held-out data or new task conditions. Good model fit is evidence that a model captures some data structure. It is not proof of model uniqueness. A serious comparison asks whether the model explains results that competitors cannot, with appropriate penalties for flexibility and explicit tests of identifiability.

Why Repetition Suppression Is Not Decisive Proof

Repetition suppression refers to decreased neural responses when a stimulus is repeated. A predictive interpretation says that repeated stimuli become expected and therefore generate less mismatch. Another interpretation emphasizes adaptation in the responding cells or networks. Both mechanisms may contribute, and specific tasks can recruit several processes simultaneously.

A strong design changes expectation while holding relevant stimulus histories as constant as possible. Omission trials, controlled sequences, matched probabilities, and model-based contrasts can reduce confounds. Yet each design introduces its own assumptions about which information the organism learned and how the response should be interpreted.

The reviews by Garrido et al. (2009) and Heilbron and Chait (2018) show why apparently simple expected-versus-unexpected contrasts are insufficient on their own.

Is Prediction the Same as Top-Down Feedback?

Top-down feedback refers broadly to influences from higher-order or later processing stages on lower-order sensory activity. Predictive information may be conveyed through feedback, but it can also arise in local circuitry, recurrent dynamics, learned temporal structure, or motor pathways. Conversely, feedback may carry attention, task goals, contextual information, or other signals without implementing a particular probabilistic prediction.

Teufel and Fletcher (2020) caution against reducing the nervous system’s various predictive processes to top-down signaling alone. Experiments need anatomical, temporal, and representational evidence before assigning a predictive role to any one connection.

Competing Explanations and Theoretical Disagreements

Adaptation and feature-selective processing

Stimulus-specific adaptation can change the size of responses when recent stimulation changes neural responsiveness. Feature-selective and feedforward systems can sometimes produce apparent expectation effects without encoding the hierarchical errors hypothesized by predictive coding. A discriminating experiment needs a prediction that separates these accounts.

Recurrent and dynamical-systems descriptions

Recurrent neural networks can use context and history to alter activity over time. Dynamical-systems models may describe transitions and stable states without explicitly labeling separate neuron classes as prediction and error carriers. Some models can be translated between descriptions; others make distinct commitments. Researchers should compare explicit models rather than contrast broad metaphors.

Efficient coding and information-theoretic alternatives

Neural systems may reduce redundancy or adapt representations to the statistical properties of inputs. Such behavior can resemble certain predictive-coding outcomes. Showing efficient representation supports adaptive neural computation, but further tests are needed to identify a particular hierarchical algorithm.

Disagreement within predictive coding itself

Spratling (2017) documents multiple predictive-coding algorithms with different generative models and optimization rules. Consequently, a result that falsifies one implementation does not refute every possible predictive approach, while a result consistent with one implementation does not establish the whole family.

The problem of explanatory flexibility

An overly broad theory can be adjusted after the fact to explain nearly any pattern, including greater activity, lower activity, or unchanged activity. A scientifically useful hypothesis specifies in advance which patterns should occur under defined parameter values and what results would count against it. Furutachi and Hofer (2026) argue for clarifying the information encoded by putative sensory prediction-error responses, which is central to reducing that flexibility.

Predictive Processing, Learning, and Attention

A predictive model may infer the present state of the world, learn more stable relationships from repeated experience, or alter which information is attended to. These processes operate at different timescales and require separate measurements. A mismatch on one trial is not automatically proof that synaptic learning occurred, and a learned expectation is not always the same thing as conscious attention.

Attention can modify perception by changing the information sampled, behavioral goals, gain, or response selection. Some predictive-coding accounts describe attention through precision weighting; other theories characterize attention differently. Their overlap produces experimentally useful questions rather than a reason to collapse their meanings.

Learning-related mechanisms and reward learning are investigated throughout neuroscience. The EN-30 article on reward prediction error covers a specific learning computation; the present article addresses broader predictive models of neural inference.

Predictive Processing and Conscious Experience

Some philosophers and neuroscientists use predictive frameworks to explain the organization of perceptual experience, its apparent stability, and the influence of expectations on what becomes salient. Such accounts generate questions about conscious access, confidence, and perceptual interpretation.

The fact that perception uses contextual information does not entail a complete theory of consciousness. Neural correlates of expected or unexpected stimulation do not settle why or whether a particular experience is subjective. Consciousness-specific explanations and experiments have a different canonical scope from the present neuroscience-model overview.

The popular phrase that the brain hallucinates reality is especially misleading. Ordinary perception is continuously constrained by sensory information, bodily action, environmental structure, and learning. Clinical hallucinations are distinct phenomena with heterogeneous mechanisms and clinical contexts. No inference about an individual’s mental health follows from the existence of prediction-dependent perception.

Can Predictive Processing Explain Psychiatric Conditions?

Researchers have proposed predictive accounts of hallucinations, anxiety, autism, functional symptoms, and other clinical phenomena. These hypotheses may help organize questions about expectations, uncertainty, attention, sensory processing, and learning. The models vary considerably, and different populations or symptom domains should not be collapsed into one explanation.

A hypothesis that altered precision weighting contributes to a symptom does not demonstrate a disorder-specific cause, much less an individual diagnostic mechanism. Observed group-level differences in task performance, electrophysiology, or imaging do not automatically generalize to each patient. A computational fit does not make predictive processing a validated clinical biomarker.

Care, diagnosis, treatment, and condition-specific neurobiology belong to the appropriate clinical evidence base, not to speculation derived from a unified theory. Established interventions require their own clinical trials and guidelines. Current predictive models are best used to formulate measurable research hypotheses, and diagnostic or therapeutic decisions should rely on clinically validated methods.

Does Predictive Processing Tell Us How Artificial Intelligence Works?

Prediction is a useful organizing word in both computational neuroscience and machine learning, yet similar vocabulary can refer to distinct architectures, data, objectives, and validation standards. Some machine-learning systems learn predictive representations; some models are explicitly inspired by biological predictive coding; many successful AI systems use other training and inference procedures.

The ability of an artificial system to predict text, images, or actions does not show that biological cortex implements the same algorithm. Nor does a successful brain-inspired algorithm establish a claim about machine consciousness or subjective experience. Neuroscience–AI comparisons are strongest when the task, learned representation, causal mechanism, and generalization test are all specified.

What Would Count as Stronger Evidence?

A decisive research program would produce preregistered predictions about representational content, neural directionality, cell classes, timing, and causal perturbations. The same task would be analyzed with competing computational models that differ on clearly stated outcomes. Experiments would separate stimulus probability from repetition history, control behavioral relevance, and use independent datasets for model evaluation.

At the circuit level, researchers could identify populations whose activity tracks quantitatively estimated prediction variables versus sensory discrepancy variables, then manipulate each population and assess the expected effects on perception. Cross-area recordings could test whether hypothesized feedback and error channels carry the predicted information at the predicted times. Cross-species work could establish which mechanisms generalize, while carefully documenting species- and task-specific constraints.

Contemporary research remains active. The 2026 Annual Review calls for sharper distinctions among superficially similar sensory error signals, and the 2024 empirical review makes clear why model selection and neural validation are essential.

What Is Established, Preliminary, and Debated?

Established evidence: prior expectations and contextual regularities influence behavioral perception and activity in multiple sensory systems. Neural circuits register unexpected events and respond to relationships among stimuli, recent history, and movement. This broad conclusion is supported by converging psychophysics and physiology.

Supported but model-dependent: particular studies show signals consistent with hierarchical expectation and mismatch mechanisms, including auditory novelty and visuomotor mismatch findings. These results are informative, but the unique neural algorithm and its scope remain matters for further tests.

Preliminary or contested: universal cortical predictive coding, a single canonical error-neuron circuit across modalities, unambiguous mappings from EEG/fMRI signatures onto formal precision weighting, and broad clinical claims about disorder mechanisms.

The strongest present conclusion is that predictive processing supplies a rigorous set of candidate models whose success depends on how precisely each one is specified and tested. The brain’s use of predictive information is well evidenced; the unity and biological generality of particular explanatory schemes are the continuing scientific debate.

Common Misconceptions About Predictive Processing

“The brain predicts every event before it happens.”

The theory concerns probabilistic estimates at particular timescales and levels of processing. The nervous system also responds to immediate inputs, novel events, and changing physiological conditions. Predictive models do not imply omniscience, conscious foresight, or successful advance prediction of every stimulus.

“Every unexpected response is a prediction-error neuron.”

A signal responsive to novelty may reflect adaptation, attention, salience, feature detection, or different learning processes. Formal error coding requires tests of the specific predicted variable and competing explanations.

“Prediction error always means dopamine.”

Dopamine is central to particular reinforcement-learning accounts, and dopamine biology includes many signals and functions. Sensory mismatch activity can be investigated in cortical circuits without treating it as a dopamine measurement. The article on dopamine and learning covers that narrower neurochemical territory.

“If the brain uses predictions, perception is fictional.”

Normal perception integrates sensory evidence with learned context. Context sensitivity does not erase contact with the environment. The accuracy of perception is assessed through behavior, signal conditions, and appropriate experimental measures.

“Predictive processing has already solved consciousness and mental illness.”

Neither subjective experience nor clinical diagnosis follows from a generic prediction model. Claims in those areas require separate, specific evidence.

“A neural model that fits the data must be true.”

Different architectures can fit the same dataset. Better evidence comes from discriminating predictions, replication, generalization, and interventions that exclude reasonable competitors.

Frequently Asked Questions

What is the simplest definition of predictive processing?

Predictive processing is the idea that the nervous system uses expectations learned from past experience and context to help interpret incoming signals and adapt when outcomes differ. Specific versions propose precise mathematical and neuronal mechanisms.

Is predictive processing a theory or a proven fact?

It is a family of research theories. Expectations demonstrably influence perception and neural responses. Specific universal claims about hierarchical cortical predictive-coding architecture remain under active evaluation.

What is the difference between predictive processing and predictive coding?

Predictive processing is the wider conceptual family. Predictive coding refers to more specific algorithms that use predictions and discrepancy signals to update representations.

Does predictive coding mean the brain ignores sensory input?

No. Sensory observations constrain estimates and can cause the model to update. Predictive accounts generally depend on the interaction of incoming evidence and model-derived expectations.

Is the brain Bayesian?

Some behavior resembles Bayesian integration under specified task conditions. Similar behavior can arise through different algorithms, and departures from ideal Bayesian predictions occur. The neural implementation must be tested separately.

What is a neural prediction error?

It is a model-defined difference or discrepancy between expected and observed information, or their representations. Its variable, direction, and biological mechanism vary across tasks.

How does surprise affect brain activity?

Unusual or rule-violating events often elicit distinct neural responses. Their causes can include adaptation, salience, learning, and prediction-related processes; larger amplitude alone does not reveal the mechanism.

What is mismatch negativity?

Mismatch negativity is an EEG/MEG response associated with deviations from an auditory regularity. Its interpretation requires distinguishing changes in learned expectations from repetition-related neural adaptation.

Can fMRI prove predictive coding?

fMRI provides indirect spatially resolved signals related to brain activity. By itself it cannot identify the exact computation performed by a neuronal population, although careful model comparisons can test predictions of specific accounts.

What is precision weighting?

It refers to weighting evidence according to estimated reliability or uncertainty within a mathematical model. Neuroscientific mappings to neural gain or attention are hypotheses requiring independent tests.

How does predictive processing relate to learning?

Predictions can be updated by evidence, and experiences may alter longer-term model parameters. Different predictive theories specify different learning mechanisms and time scales.

Is predictive processing the same as reward prediction error?

No. Reward prediction error concerns discrepancies between expected and actual reward outcomes in specific learning contexts. Predictive-processing models also address sensory and state-estimation computations.

Does predictive processing explain anxiety, autism, or psychosis?

It motivates research hypotheses about particular cognitive or sensory processes. It does not by itself establish a unique cause, diagnosis, biomarker, or proven treatment for any condition.

Is active inference the same as predictive coding?

Active inference includes model-based action and policy selection. Predictive coding usually refers more narrowly to how predictions and discrepancies may be represented and updated during inference.

What is the biggest unresolved issue?

Researchers must identify which proposed neural signals truly encode predictions or prediction errors, distinguish them from alternative computations, and demonstrate whether specific architectures generalize across species, circuits, and tasks.

References

Aitchison, L., & Lengyel, M. (2017). With or without you: Predictive coding and Bayesian inference in the brain. Current Opinion in Neurobiology, 46, 219–227. https://doi.org/10.1016/j.conb.2017.08.010

Bastos, A. M., et al. (2012). Canonical microcircuits for predictive coding. Neuron, 76(4), 695–711. https://doi.org/10.1016/j.neuron.2012.10.038

de Lange, F. P., Heilbron, M., & Kok, P. (2018). How do expectations shape perception? Trends in Cognitive Sciences, 22(9), 764–779. https://doi.org/10.1016/j.tics.2018.06.002

Friston, K. (2010). The free-energy principle: A unified brain theory? Nature Reviews Neuroscience, 11, 127–138. https://doi.org/10.1038/nrn2787

Furutachi, S., & Hofer, S. B. (2026). Rethinking predictive processing. Annual Review of Neuroscience, 49, 471–494. https://doi.org/10.1146/annurev-neuro-102124-031410

Garrido, M. I., Kilner, J. M., Stephan, K. E., & Friston, K. J. (2009). The mismatch negativity: A review of underlying mechanisms. Clinical Neurophysiology, 120(3), 453–463. https://doi.org/10.1016/j.clinph.2008.11.029

Heilbron, M., & Chait, M. (2018). Great expectations: Is there evidence for predictive coding in auditory cortex? Neuroscience, 389, 54–73. https://doi.org/10.1016/j.neuroscience.2017.07.061

Hodson, R., Mehta, M., & Smith, R. (2024). The empirical status of predictive coding and active inference. Neuroscience & Biobehavioral Reviews, 157, 105473. https://doi.org/10.1016/j.neubiorev.2023.105473

Keller, G. B., Bonhoeffer, T., & Hübener, M. (2012). Sensorimotor mismatch signals in primary visual cortex of the behaving mouse. Neuron, 74(5), 809–815. https://doi.org/10.1016/j.neuron.2012.03.040

Keller, G. B., & Mrsic-Flogel, T. D. (2018). Predictive processing: A canonical cortical computation. Neuron, 100(2), 424–435. https://doi.org/10.1016/j.neuron.2018.10.003

Kok, P., Jehee, J. F. M., & de Lange, F. P. (2012). Less is more: Expectation sharpens representations in the primary visual cortex. Neuron, 75(2), 265–270. https://doi.org/10.1016/j.neuron.2012.04.034

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://doi.org/10.1038/4580

Spratling, M. W. (2017). A review of predictive coding algorithms. Brain and Cognition, 112, 92–97. https://doi.org/10.1016/j.bandc.2015.11.003

Teufel, C., & Fletcher, P. C. (2020). Forms of prediction in the nervous system. Nature Reviews Neuroscience, 21, 231–242. https://doi.org/10.1038/s41583-020-0275-5

Wacongne, C., et al. (2011). Evidence for a hierarchy of predictions and prediction errors in human cortex. Proceedings of the National Academy of Sciences, 108(51), 20754–20759. https://doi.org/10.1073/pnas.1117807108

Walsh, K. S., McGovern, D. P., Clark, A., & O’Connell, R. G. (2020). Evaluating the neurophysiological evidence for predictive processing as a model of perception. Annals of the New York Academy of Sciences, 1464(1), 242–268. https://doi.org/10.1111/nyas.14321

Psychological Encyclopedia