Predictive Processing in Perception: Theory, Evidence, and Debate

Author: Angela Bogdanova · Published: October 16, 2026 · Editorial Policy

Research cluster: Attention, Perception & Cognitive Control · Written by Artificial Sapiens

Predictive processing in perception is a family of psychological and computational theories proposing that perceptual systems interpret uncertain sensory information in light of expectations about the world. In these accounts, perception involves estimating what most likely caused the signals arriving at the senses, taking account of prior experience, the present context, and the reliability of available evidence. Some versions go further and propose that differences between predicted and actual signals—prediction errors—drive continual revision of these estimates. The central scientific question is how far this framework explains what people actually see, hear, and feel, and which of its proposed mechanisms have survived discriminating tests. de Lange, Heilbron, and Kok (2018); Hodson, Mehta, and Smith (2024).

The evidence supports a substantial role for learned regularities, expectation, and uncertainty in perception. Carefully controlled psychophysical experiments demonstrate systematic changes in perceptual judgments when the distribution or reliability of information changes. Some neuroimaging results show stimulus-specific anticipatory representations and expectation-related changes in sensory activity. The broader claim that perception everywhere depends on a single predictive-coding circuit remains a theoretical proposal requiring further tests. Distinguishing these conclusions is essential to understanding both the success and the controversy surrounding predictive processing. Knill and Pouget (2004); Feuerriegel, Vogels, and Kovács (2021).

What Does Predictive Processing Mean in Perceptual Psychology?

The term describes explanations in which an observer uses information about the likely structure of the environment to interpret sensory evidence. A sensory signal is often compatible with several possible events. The size and shape of a retinal image, for instance, depend jointly on the object, viewpoint, illumination, and surrounding scene. Perception must settle on a useful interpretation despite this underdetermination. Models of perceptual inference explain why an observer may combine the current image with statistical knowledge of how objects, surfaces, or movements typically behave. Kersten, Mamassian, and Yuille (2004).

A perceptual prediction is an expectation formulated by a model or an observer about a signal, feature, or hidden cause. It need not be an explicit forecast or a consciously entertained belief. The relevant information might arise from a learned visual sequence, a cue indicating which orientation is probable, familiarity with an object, or the characteristic distribution of motion in a particular task. Its influence can occur before stimulus presentation, during signal processing, or while a person selects a response.

Psychologists use the term at several levels. A functional claim states that useful perception depends on exploiting regularities. A computational model specifies probabilities and estimates. An algorithm specifies how representations are updated. A biological account proposes particular neural populations and connections. Evidence at one level constrains but does not automatically prove the others. This distinction prevents the observation of an expectation effect from being mistaken for direct validation of a specific cortical circuit. Hodson et al. (2024).

Why sensory information requires interpretation

Images, sounds, and touches provide measurements rather than direct, complete descriptions of the world. The same retinal contour may correspond to different three-dimensional shapes; a weak sound may be speech or background noise; a moving pattern may be compatible with several velocities. An effective observer must manage ambiguities, measurement noise, and changing environmental probabilities. Predictive processing offers one formal vocabulary for explaining this process, while theories of perceptual organization, signal detection, and sensory physiology specify complementary constraints.

The result is a continuously revised perceptual estimate. When the available evidence is clear, the estimate may rely mainly on the signal. When evidence is weak or ambiguous, learned regularities can contribute more. In a quantitative model, this dependence can be represented and tested rather than inferred merely from familiar everyday examples. Knill and Pouget (2004).

Historical Foundations: Inference, Probability, and Predictive Coding

The idea that perception involves interpretation precedes modern predictive-processing terminology. Research on perceptual organization and the effects of context showed that what an observer reports depends on relationships among stimuli, not simply isolated sensory elements. Probabilistic accounts subsequently gave these insights precise mathematical forms, asking how prior knowledge and measurement uncertainty could determine the best estimate of an object or event.

A particularly influential computational development was Bayesian perception. Reviews by Kersten, Mamassian, and Yuille (2004) and Knill and Pouget (2004) explained how sensory likelihoods and learned prior distributions might be combined to estimate hidden causes. This program generated quantitative predictions about cue integration, visual motion, and context-dependent judgments.

Predictive coding added more specific proposals about how such estimates could be implemented. Rao and Ballard (1999) modeled a visual hierarchy in which descending signals predicted activity at lower levels and ascending signals represented residual differences. Their simulations showed that particular receptive-field phenomena could emerge from this organization. As a model, this was a powerful proof of possibility; the model's success did not by itself establish that the visual system uses exactly the simulated architecture.

More recent work has integrated Bayesian observer models, hierarchical generative models, sensory prediction error, precision, and action-dependent sampling under the broader label of predictive processing. These terms should remain analytically distinct because they make different empirical commitments. A theory that fits response variability need not specify error-encoding neurons, and a recorded neural mismatch signal need not reveal the computations producing a perceptual judgment. Hodson et al. (2024); Furutachi and Hofer (2026).

The Bayesian Core: Priors, Likelihoods, and Posterior Estimates

Bayesian models provide the clearest formal introduction to perceptual prediction. Suppose an observer wants to estimate a hidden property H, such as an object's speed, from a sensory measurement S. The model combines a prior distribution P(H), describing how plausible different values are before considering the current measurement, with a likelihood P(S | H), describing how likely that measurement would be under each possible value. Bayes' rule states that P(H | S) is proportional to P(S | H) multiplied by P(H). The posterior distribution represents the model's updated uncertainty after the measurement.

A prior is an explicit component of a model, inferred from an experimentally manipulated distribution or estimated from responses under assumptions. It is not necessarily a personal belief, a memory that someone can report, or a fixed preference. A likelihood characterizes the statistical relationship between possible stimulus properties and noisy measurements. A posterior is a distribution rather than automatically the single experience a participant reports. Researchers must also specify a decision rule—such as the posterior mean, mode, or an estimate minimizing expected loss—to predict a behavioral answer.

A concrete example: estimating speed in poor visibility

Imagine judging the speed of a faint moving pattern. If sensory measurements are noisy and slower movements are common within the observer's learned environment, a Bayesian model can predict estimates shifted toward lower speeds. When visual contrast improves, a more reliable measurement has greater influence. The same physical velocity can therefore appear different across contrast conditions. This prediction was formalized in the motion-perception work of Weiss, Simoncelli, and Adelson (2002) and quantitatively explored by Stocker and Simoncelli (2006).

That research is a case of quantitative psychophysics: investigators measure biases and variability across controlled stimuli and compare the resulting response distributions with model predictions. It supports the usefulness of probabilistic perceptual estimation. It does not establish that individual neurons encode literal probability tables or that all visual phenomena use the same prior distribution.

Reliability and precision

In many mathematical formulations, greater certainty gives an observation or prior greater weight. Precision is commonly defined as inverse variance within a Gaussian model: a narrower uncertainty distribution has higher precision. This helps explain why reliable visual information can dominate a noisy cue in one setting while touch or another signal becomes influential when visual information is poor. The precise weighting rule depends on the measurement model, the correlations between cues, and the observer's task.

Ernst and Banks (2002) experimentally measured visual and haptic estimates and found integrated behavior close to a variance-minimizing model in a particular height-judgment task. This result provides a rigorous example of uncertainty-sensitive sensory integration. It should be described as support for an observer-level computation, not as proof of a specific hierarchical error-message architecture. The distinction matters because several algorithms can produce similar weighted estimates.

Estimates, perception, and response decisions

A participant's key press or verbal report is the output of perception combined with task comprehension, response criteria, and possible motor or decisional biases. A model that reproduces the distribution of responses may be a good description of observed behavior even when the internal perceptual representation differs from its assumptions. Strong tests separately measure discriminability, biases, confidence, reaction time, and changes caused by experimental manipulation whenever those measurements can distinguish competing processes.

Predictive Coding: What Is Supposed to Be Predicted and Updated?

Predictive coding is a more specific hypothesis within the predictive-processing family. Many formulations posit a hierarchy that generates expectations about activity at a lower level and computes residual differences when new sensory data arrive. Those differences can guide the revision of a representation, adjustment of model parameters, or a change in confidence. In a simple illustrative case, the model anticipates an edge at one location; the input differs, so the discrepancy changes the inferred object position.

The word error is technical. Prediction error is a difference between modeled expectation and observation, often weighted by uncertainty; it is not a mistake, a clinical symptom, or a conscious experience of surprise. An unexpected event may produce both a measurable neural response and a noticeable perceptual change, but those two outcomes can be generated by different processes. Behavioral surprise, subjective novelty, statistical improbability, and a modeled error term should not be treated as interchangeable variables.

An influential proposal assigns descending predictions and ascending error signals to different pathways, while later versions include gain adjustment according to precision. Rao and Ballard (1999) provide a foundational computational model. Contemporary reviews emphasize variation among implementations and the challenge of finding experimental signatures that uniquely identify the proposed algorithm. Hodson et al. (2024); Furutachi and Hofer (2026).

Predictive processing versus predictive coding

Predictive processing is the broad claim that predictions and uncertainty help organize perception. Predictive coding specifies mechanisms that represent and reduce residual discrepancies. Bayesian perception formally describes the combination of evidence and prior information, but does not select a unique implementation. These propositions can be separately supported or refuted. A statistical observer model may work extremely well while a proposed neural-coding rule fails.

Predictive processing versus active inference

Active inference extends related ideas to action and information-seeking by modeling how an organism may select actions that change or sample its sensory inputs. Its explanatory scope includes policy selection and control, which require different behavioral tests from passive visual judgments. Hodson et al. (2024) reviewed predictive coding and active inference separately and found that broad model fit alone leaves important validation questions open. The present article centers on perceptual inference; action and neural implementation are discussed only where they illuminate perception.

What Do Behavioral Experiments Actually Show?

Psychophysical research is particularly informative because an investigator can independently manipulate what is shown, what is statistically likely, and how accurately a person reports it. Predictions become testable when a model specifies which change in probability or uncertainty should shift a judgment, in what direction, and by how much. The strongest designs measure whole response distributions or use conditions that distinguish perceptual and decisional effects.

Motion and the role of prior distributions

Weiss et al. (2002) showed how a slow-motion prior in a noisy-estimation model could generate several motion illusions. Stocker and Simoncelli (2006) combined discrimination measurements with model fitting to infer internal noise and a prior distribution across motion speeds. These results support the idea that visual estimates can reflect both the available evidence and a learned or assumed distribution of likely causes. The particular prior and noise model must still be justified rather than chosen solely to fit the final plot.

Combining information from sight and touch

A participant estimating object height can receive separate visual and tactile measurements. Ernst and Banks (2002) found that their participants' integration closely resembled a statistically efficient combination of the two signals under the measured conditions. This result is stronger than an anecdote about expectations because the model was constrained by the reliability of the separate sources. Yet cue integration is one computational problem among many: the outcome does not imply that every sensory estimate uses hierarchical prediction error.

Context and timing

Predictive effects extend to estimates of duration. Jazayeri and Shadlen (2010) asked people to reproduce intervals drawn from experimentally specified distributions. Responses systematically regressed toward the mean of the sampled distribution, a pattern captured by a Bayesian model taking measurement variability and temporal context into account. The implication is that an observer can adapt estimates to the statistics of the task. The broader mechanisms of time perception deserve separate treatment, but this study illustrates a general methodological strength: the prior can be manipulated rather than merely presumed.

The limits of an ideal-observer fit

A good Bayesian fit establishes that a specified probabilistic model describes the measured response pattern under its assumptions. It is not a direct recording of the computation a nervous system performs. Competing models may approximate the same curves with different latent variables, learning rules, or response strategies. Strong inference therefore requires predictions made before examining the results, parameter recovery, comparison against plausible alternatives, and tests on held-out tasks or conditions.

Can Expectations Change What a Person Perceives?

Many experiments demonstrate that expected and unexpected events produce different performance or neural signals. The interpretation of those differences depends on whether investigators have separated changes in perceptual sensitivity from changes in response preference. An observer who expects a faint stimulus may be more willing to report that it occurred. That change could shift a detection rate without changing the sensory representation itself. Signal-detection methods help distinguish such scenarios when their assumptions are met.

Expectation can also alter the quality or timing of sensory representations. In human neuroimaging, Kok, Jehee, and de Lange (2012) found that expected stimuli elicited lower average responses in primary visual cortex while their multivariate representations were sharper, with relationships to behavioral benefits. The combination cautions against equating a lower signal amplitude with weaker perceptual information. Kok, Mostert, and de Lange (2017) reported activity patterns consistent with expected stimulus templates before stimulus presentation, providing a more temporally specific test of anticipatory sensory representation.

Both findings constrain models of the interaction between prior expectations and sensory processing. Yet a representation consistent with an expected feature does not independently identify a universal prediction-error subtraction operation. Task demands, attentional selection, learning, and response preparation can influence similar measures. Investigators must demonstrate why a particular result discriminates one model rather than merely agreeing with it. de Lange et al. (2018).

Expectations and perceptual attention are different variables

Expectation describes a probability concerning an event; attention concerns the prioritization of information for the observer's goals and task. A highly probable object need not be relevant to the task, while a rare target can command attention precisely because it matters. These variables often interact, but they should be manipulated independently. Summerfield and Egner (2009) developed this distinction in the study of visual cognition.

Some prediction effects have been interpreted as sensory attenuation, sharpening, improved efficiency, or decision bias. All are possible outcomes of different experimental conditions. An expectation effect needs its own dependent measure and a discriminating causal design; the broad concept alone does not determine whether the person sees more strongly, responds faster, reports more confidently, or shows less average neural activity.

The Perceptual Prediction Paradox

One of the deepest debates concerns the apparent conflict between two plausible claims. If prior expectations support accurate interpretation of noisy input, people may tend to experience expected objects or features more readily. If prediction errors are prioritized because unexpected signals carry information for updating a model, unexpected inputs may receive greater processing. These predictions can appear to point in opposite directions.

Press, Kok, and Yon (2020) called attention to this perceptual prediction paradox. Their analysis separates the objective of generating accurate representations from the objective of learning something new. They proposed that different processes or temporal stages may help explain why expectation can facilitate an initial interpretation while surprising evidence becomes more influential later. The proposal is a substantive theoretical synthesis whose precise conditions require empirical testing.

Accuracy and information gain

An estimate can be statistically accurate because it uses reliable priors, even if it contains little new information relative to what was already expected. Conversely, an unexpected observation may be informative for updating a model even if its initial interpretation is uncertain. Research should specify which outcome is being measured. Better discrimination of a likely signal and a larger neural response to a rare signal are not logically inconsistent, since they may concern different computations and timescales.

When does prediction lead to facilitation or attenuation?

The answer depends on the operational definition. Facilitation might mean faster identification, more accurate discrimination, or stronger perceptual appearance. Attenuation may mean reduced neural amplitude or diminished reported intensity. Suppression in a brain measurement is not automatically perceptual suppression. The competing hypotheses become testable only after the dependent variable, learning period, stimulus statistics, and decision rule have been specified. Press et al. (2020).

Evidence for Expectation Suppression: What Survives the Controls?

Expectation suppression is the proposal that anticipated sensory events elicit smaller neural responses than unanticipated ones, potentially because part of the input has been predicted. It has served as an influential test case for predictive-coding theories, but several other processes can cause reduced responses. Repeating a stimulus can induce adaptation; directing attention elsewhere can modify measured activity; an oddball event may be novel or surprising; and common versus rare sequences differ in their statistical histories.

Feuerriegel, Vogels, and Kovács (2021) reviewed expectation-suppression evidence across visual paradigms and identified these substantial confounds. They found support in a restricted subset of designs involving extended statistical learning, while evidence was inconsistent in many paradigms involving probabilities learned during brief sessions. This review narrows the conditions under which an observation can count as compelling support for genuine expectation suppression.

The general lesson is methodological. A researcher must vary expectation independently of repetition, attention, and physical stimulus properties. When a purported signature disappears after these controls, that finding constrains the theory. When it remains, the next question is whether it distinguishes an error-coding algorithm from competing mechanisms of adaptation, gain control, or associative learning. Feuerriegel et al. (2021).

Repetition is not the same as prediction

A repeated image is more familiar and often more likely under a sequence than a novel one, but repetition and conditional expectancy can diverge. Responses can decline because the same neural population was recently activated, even when there is no learned prediction about which particular stimulus should appear next. A clean test uses experimental sequences that distinguish repeating a physical event from expecting one. Reviews of repetition suppression underline that multiple sub-processes and timescales are involved. Grotheer and Kovács (2016).

How Predictive-Processing Theories Are Measured and Tested

The credibility of an explanatory model depends on designs that establish what the data can and cannot identify. Laboratory researchers combine psychophysical tasks, statistical modeling, neuroimaging, electrophysiology, and in some cases causal intervention. Each method answers a different question. A good research program specifies the inferential gap between the observable measurement and the latent mechanism under consideration.

Psychophysics and signal-detection analysis

Forced-choice discrimination tasks can estimate sensitivity to differences in orientation, movement, timing, or sensory intensity. Detection tasks can separate changes in sensitivity and response criterion under a signal-detection model. Trial-by-trial responses reveal whether a prior affects estimated stimulus magnitude, the willingness to report a target, response time, or confidence. Quantifying these effects avoids treating every increase in detection as a gain in perceptual sensitivity.

Controlled priors and sensory likelihoods

An ideal design experimentally varies a cue's reliability or the distribution of stimuli over trials while holding actual inputs comparable at critical test points. Changes in sensory noise, such as lower contrast, provide an independent test of the prediction that priors matter more when likelihoods are broad. Research on speed and interval estimation demonstrates the value of measuring complete behavioral patterns, including bias and variability. Stocker and Simoncelli (2006); Jazayeri and Shadlen (2010).

Model comparison and identifiability

Researchers should specify models before evaluating them, estimate parameters with attention to recoverability, and compare predicted results on data not used to fit the models. A Bayesian observer model, a reinforcement or associative learner, and a response-bias account may sometimes fit similar averages. Discriminating manipulations include unexpected reversals of statistical contingencies, changes in cue validity, novel conditions, and measures of learning over time. Conclusions should concern the family of alternatives actually tested, with uncertainty reported at the level justified by the design.

Neurophysiological and imaging measurements

Functional MRI, EEG, MEG, and electrophysiological recordings can reveal anticipatory responses, changes in stimulus representations, or differential reactions to mismatches. Their signals are indirect with respect to the model variables of perceptual probability and precision. Average activation, decoded information, timing, and causal necessity must be interpreted separately. A neurophysiological signature is stronger evidence for a mechanistic theory when the experiment specifies what alternative processes would predict and can rule them out. Furutachi and Hofer (2026).

Replication, generalization, and ecological validity

Many experiments use simple, repeated patterns over minutes or hours. Everyday perception concerns rich scenes, rapidly changing goals, multiple senses, and long learning histories. A finding that survives careful laboratory control may still generalize differently to naturalistic contexts. Stronger programs replicate with adequate sample sizes, share data and analysis code, test diverse participant groups where the question warrants it, and validate the model across tasks rather than one convenient dataset.

Criticisms, Alternative Accounts, and Open Scientific Questions

Debate about predictive processing is productive when it concerns specific, falsifiable claims. Broad statements that brains use expectations are compatible with extensive existing research, while a claim that every cortical stage implements the same precision-weighted error algorithm is far more restrictive. The question is which explanatory commitments belong to a particular model and what evidence would count against them. Hodson et al. (2024).

Can multiple theories explain the same results?

Yes. Neural response reduction may be explained through sensory adaptation, fatigue, local inhibition, or changed attention. Response facilitation may reflect an altered decision criterion or better allocation of processing resources. These alternatives need explicit, matched tests; the presence of an expectation effect does not select predictive coding by itself. Feuerriegel et al. (2021).

Does Bayesian behavior mean the brain literally computes Bayes' rule?

A computational model can describe observable performance closely without identifying its implementation. An observer may approximate probabilistic integration through learned weights, heuristic rules, neural population codes, or other mechanisms. The mathematical optimum can serve as a benchmark, but successful benchmark performance alone does not determine the algorithm. Knill and Pouget (2004); Loued-Khenissi and Pascucci (2026).

Are error signals uniquely identified?

Not usually by one response signature. An unexpectedly strong neural reaction could reflect an error computation, attention, novelty, learned salience, or a prediction-dependent state change. Furutachi and Hofer (2026) argue that resolving what these signals represent is central to evaluating predictive processing. Experiments need contrasts that isolate the modeled information content and a causal account of how the signals change subsequent estimates.

Can a broad framework become difficult to falsify?

A theory loses empirical precision if every neural increase, decrease, behavioral bias, and failed prediction can be explained after the fact by adjusting unspecified priors or precision weights. Stronger versions constrain these quantities independently, state which patterns cannot occur under the model, and predefine meaningful alternatives. The empirical review by Hodson et al. (2024) concludes that support for specific predictive-coding hypotheses remains modest, underscoring the importance of discriminating research rather than broad conceptual agreement.

Everyday Perception: Useful Implications and Their Limits

Predictive models help explain why context can make ambiguous sensory data easier to interpret and why unusual conditions can produce systematic perceptual errors. An observer who expects a certain word in a noisy sentence may use that context to resolve uncertainty. A person perceiving motion in low contrast may exhibit an estimate shaped by the statistical assumptions of the task. These examples illustrate computational principles, but exact effects depend on the measured modality, stimulus statistics, and experimental setup. de Lange et al. (2018); Stocker and Simoncelli (2006).

The framework also highlights why confidence and accuracy can diverge. An estimate can feel compelling because prior information strongly favors it while the actual sensory evidence is weak. How this affects a particular person's judgments is a question for behavioral measurement. The model does not provide a universal method for diagnosing individual perceptual style or correcting daily attention problems.

Some popular descriptions claim that perception is merely hallucination or that prediction explains every illusion. Such language obscures the role of the physical environment, sensory evidence, learning, and corrective feedback. Ordinary perceptual inference remains constrained by incoming information. Clinical hallucinations are diverse clinical phenomena assessed through separate evidence and criteria; they cannot be inferred from ordinary prior-weighted perception. Kersten et al. (2004).

Predictive processing is also used in discussions of consciousness, psychiatric disorders, learning, and artificial systems. Those applications have their own explanatory burdens. An observer model of uncertain visual input does not establish a theory of subjective experience, a medical diagnosis, or a general guarantee of intervention effectiveness. The scientific use of the model depends on keeping its measurable claims tied to the domain in which they were tested.

What Is Established, What Remains Provisional, and What Is Contested?

Established: Context and learned regularities can systematically influence perceptual judgments. The reliability of sensory evidence affects cue integration and estimates in well-specified tasks. Multiple experiments demonstrate expectation-related changes in sensory activity, and probabilistic models predict important patterns of psychophysical behavior. Ernst and Banks (2002); Stocker and Simoncelli (2006); de Lange et al. (2018).

Supported under specified conditions: Stimulus expectations can sharpen representations, and some experimental paradigms find effects consistent with sensory prediction or suppression. Different measurement methods identify partly distinct components of such effects. The strength of support depends on whether the design separates prior probability, attention, adaptation, novelty, and the final response decision. Kok et al. (2012); Feuerriegel et al. (2021).

Limited or contested: A universal cortical predictive-coding architecture, the uniquely established identity of particular sensory error neurons, one mechanism underlying all expectation effects, and a direct inference from Bayesian model fit to literal neural probability computation. Reviews published in 2024 and 2026 emphasize these unresolved issues. Hodson et al. (2024); Furutachi and Hofer (2026).

The most defensible synthesis is that perceptual inference is a fruitful quantitative research program whose models become more convincing when their predictions are precise, their alternatives are tested, and their explanations generalize. Predictive processing matters scientifically because it offers ways to connect uncertainty, context, perceptual reports, and neural measurements—not because any single implementation has settled the whole field.

Frequently Asked Questions

Is predictive processing a proven theory of perception?

The broad influence of expectations and uncertainty on perception has substantial empirical support. Predictive processing, however, names a family of theories, and the stronger claim that one predictive-coding algorithm is the universal mechanism of perception remains under evaluation. Hodson et al. (2024).

How is predictive processing different from top-down processing?

Top-down processing is a descriptive label for the influence of higher-level knowledge or context on sensory interpretation. Predictive processing offers computational explanations of how that influence might work, often specifying probabilities, expectations, error signals, and updating. The existence of top-down effects does not select a unique computational mechanism.

Does predictive coding make perception inaccurate?

Using priors can improve accuracy when they reflect reliable environmental regularities, especially when sensory input is weak. It can also produce systematic biases under unusual or misleading conditions. Whether an estimate is accurate is an empirical property of the task and environment rather than an inevitable consequence of prediction. Weiss et al. (2002).

What is a prediction error in perception?

It is a modeled discrepancy between an expected input and an actual measurement, sometimes scaled by the estimated reliability of that discrepancy. Neural activity responding to unexpected events may be consistent with this construct, but further tests are needed to show what computation the response actually performs. Furutachi and Hofer (2026).

Is attention the same as prediction?

No. Expectation represents likelihood, while attention prioritizes information according to task demands and goals. They often interact and can produce overlapping measures, which is why factorial experimental designs and appropriate controls matter. Summerfield and Egner (2009).

Why do some experiments find reduced responses to expected stimuli?

Such reductions may reflect genuine expectancy effects, repetition, adaptation, attention, or other factors. Their interpretation depends on the experimental control conditions. Reduced measured amplitude also does not necessarily mean poorer stimulus information or weaker conscious perception. Feuerriegel et al. (2021); Kok et al. (2012).

Can a Bayesian model tell us exactly how the brain works?

It can show whether responses are consistent with a particular quantitative integration of evidence and prior probability. Identifying neural representations, algorithms, and causal circuits requires additional independent evidence. Good descriptive fits are a starting point for mechanistic research rather than its endpoint.

Related Articles

For the complementary neurological analysis of candidate circuits and neural prediction errors, see Predictive Processing in the Brain: Models, Evidence, and Debate. For the broader behavioral science of how sensory input becomes an object or scene, see Visual Perception: How We Recognize Objects and Scenes and Sensation vs. Perception: Differences and Everyday Examples.

References

• 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

• Ernst, M. O., & Banks, M. S. (2002). Humans integrate visual and haptic information in a statistically optimal fashion. Nature, 415, 429–433. https://doi.org/10.1038/415429a

• Feuerriegel, D., Vogels, R., & Kovács, G. (2021). Evaluating the evidence for expectation suppression in the visual system. Neuroscience & Biobehavioral Reviews, 126, 368–381. https://doi.org/10.1016/j.neubiorev.2021.04.002

• 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

• Grotheer, M., & Kovács, G. (2016). Can predictive coding explain repetition suppression? Cortex, 80, 113–124. https://doi.org/10.1016/j.cortex.2015.11.027

• 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

• Jazayeri, M., & Shadlen, M. N. (2010). Temporal context calibrates interval timing. Nature Neuroscience, 13, 1020–1026. https://doi.org/10.1038/nn.2590

• Kersten, D., Mamassian, P., & Yuille, A. (2004). Object perception as Bayesian inference. Annual Review of Psychology, 55, 271–304. https://doi.org/10.1146/annurev.psych.55.090902.142005

• Knill, D. C., & Pouget, A. (2004). The Bayesian brain: The role of uncertainty in neural coding and computation. Trends in Neurosciences, 27(12), 712–719. https://doi.org/10.1016/j.tins.2004.10.007

• 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

• Kok, P., Mostert, P., & de Lange, F. P. (2017). Prior expectations induce prestimulus sensory templates. Proceedings of the National Academy of Sciences, 114(39), 10473–10478. https://doi.org/10.1073/pnas.1705652114

• Loued-Khenissi, L., & Pascucci, D. (2026). Probabilistic processing: Possible, probable, or preposterous? A dialectical essay. New Ideas in Psychology, 80, 101212. https://doi.org/10.1016/j.newideapsych.2025.101212

• Press, C., Kok, P., & Yon, D. (2020). The perceptual prediction paradox. Trends in Cognitive Sciences, 24(1), 13–24. https://doi.org/10.1016/j.tics.2019.11.003

• 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

• Stocker, A. A., & Simoncelli, E. P. (2006). Noise characteristics and prior expectations in human visual speed perception. Nature Neuroscience, 9, 578–585. https://doi.org/10.1038/nn1669

• Summerfield, C., & Egner, T. (2009). Expectation (and attention) in visual cognition. Trends in Cognitive Sciences, 13(9), 403–409. https://doi.org/10.1016/j.tics.2009.06.003

• Weiss, Y., Simoncelli, E. P., & Adelson, E. H. (2002). Motion illusions as optimal percepts. Nature Neuroscience, 5, 598–604. https://doi.org/10.1038/nn0602-858