RESEARCH DOSSIER
Bayesian Perceptual Inference and Predictive Processing
Abstract. Bayesian perceptual inference explains how an observer estimates environmental causes by combining uncertain sensory measurements with learned or assumed prior distributions. Predictive processing extends this research toward candidate mechanisms in which recurrent predictions and mismatches update internal representations. This research monograph critically integrates original psychophysical experiments, formal observer models, multisensory causal-inference studies, cortical imaging and recording, neural circuit hypotheses, and theoretical reviews through 2026, including causal thalamocortical experiments and conflicting models of sensory feedback. The analysis traces the historical emergence of probabilistic perception, evaluates reliability-weighted cue integration, and distinguishes computational descriptions from neural implementation. It compares predictive coding with adaptation, attention, sampling, and competing recurrent algorithms; examines uncertainty, learning, ecological transfer, and sensorimotor decisions; and identifies model identifiability and neural causal testing as continuing challenges. Conditional Bayesian-style integration has substantial experimental support in selected perceptual tasks, while specific claims about universal cortical prediction-error mechanisms require more discriminating evidence. The concluding synthesis specifies empirical questions relevant to both biological and artificial perceptual systems.
Keywords: Bayesian perceptual inference; predictive processing; predictive coding; Bayesian brain; perceptual uncertainty; sensory likelihood; prior probability; prediction error; precision weighting; multisensory integration; Bayesian causal inference; neural sampling; visual perception; cortical feedback; computational psychology
Scientific Scope and Central Questions
Bayesian perceptual inference studies how observers estimate causes of sensory signals when the evidence is incomplete, variable, and compatible with several states of the environment. A perceptual estimate is therefore analyzed as a conditional judgment about a latent property such as direction, distance, object identity, or common source. Bayesian theory specifies how knowledge about the frequency of environmental causes, expressed as a prior distribution, can be combined with a likelihood describing possible sensory measurements. Predictive processing addresses a further question: how such estimates might be updated by recurrent systems that generate predictions and register mismatches. These research programs intersect without being logically identical. This monograph examines their conceptual connection, empirical reach, and neural implementation, concentrating on sensory perception and perceptually guided behavior rather than treating all cognition as one undifferentiated inference problem Knill & Pouget (2004) de Lange et al. (2018).
The primary scientific questions concern the conditions under which observers weight prior information according to sensory reliability, the circumstances in which a prior induces a measurable bias, the computational requirements for distinguishing a shared cause from separate causes, and the neural mechanisms through which predictions influence sensory representation. An additional question concerns model identifiability. Multiple algorithms can produce similar choices, response times, or aggregate brain signals even when they represent uncertainty differently. A good empirical explanation must consequently link independent manipulations to distinctive patterns of behavior and physiology rather than identify the word prediction with any observed repetition effect. The distinction between behavioral optimality and a confirmed neural algorithm organizes the analysis Aitchison & Lengyel (2017) Walsh et al. (2020).
The evidence considered here comprises original psychophysical experiments, computational derivations, electrophysiological and imaging studies, and critical reviews; the synthesis is a research monograph, not a protocol-driven systematic review or meta-analysis. Human perceptual tasks supply the most direct evidence about estimates and reports, while cortical recordings constrain hypotheses about implementation. Model simulations provide existence demonstrations when a proposed mechanism can reproduce phenomena, yet fitting the same data does not independently establish the mechanism. The study also recognizes that biological sensory systems learn environmental statistics and that behavior depends on goals, costs, and action. Those wider connections become relevant only where they clarify the inference process or establish testable cross-domain hypotheses Fiser et al. (2010) Haefner et al. (2016).
Historical Foundations of Probabilistic Perception
Contemporary Bayesian perception joins an older problem of inverse interpretation to the mathematics of uncertain evidence. Physical environments generate signals at sensory receptors, but the mapping from world to measurement can be many-to-one. A retinal image, for example, is compatible with different object sizes, distances, illuminations, and surfaces. The inverse problem is to infer the environmental cause given the available sensory data. Statistical estimation theory made this problem precise through likelihood functions and decision rules. Modern psychological formulations place the observer's learned or assumed regularities inside the same formal account as measurement noise. Historical foundations are best understood as an accumulation of mathematical tools and perceptual experiments rather than as a single discovery that perception always operates by explicit probability calculation Kersten et al. (2004) Knill & Pouget (2004).
The quantitative research program gained particular traction when theories generated counterintuitive predictions about perceptual errors. Rather than regarding every illusion as an arbitrary failure, an ideal-observer approach asks whether the reported estimate would be reasonable under specified prior assumptions and measurement uncertainty. Weiss, Simoncelli, and Adelson modeled visual motion using an assumption that slow velocities are more probable than rapid ones and that local measurements are noisy. Their 2002 account reproduced multiple motion phenomena within that computational family. It established that a biased percept can be compatible with a successful estimator under its own assumptions; it did not show that every visual illusion arises through that same prior or the same neural procedure Weiss et al. (2002).
A parallel history concerns cortical prediction. Rao and Ballard proposed a hierarchical visual model in which higher levels generate predictions of lower-level activity and residual errors move forward. Training on natural images produced model representations reminiscent of certain receptive-field properties. Friston subsequently developed a family of generative and variational accounts connecting cortical responses to inferred sensory causes, while later microcircuit proposals specified candidate roles for different pathways and populations. This trajectory shifted predictive coding from a computational explanation of receptive-field phenomena toward hypotheses about cortical organization. Its scientific strength depends on whether neural data discriminate its proposed signaling arrangement from alternatives that also use recurrent excitation, inhibition, or efficient coding Rao & Ballard (1999) Friston (2005) Bastos et al. (2012).
The Bayesian Observer and Its Variables
For an environmental state s and a sensory measurement m, Bayes' rule states p(s|m) = p(m|s)p(s)/p(m). The likelihood p(m|s) describes expected measurements under a hypothesized state; the prior p(s) describes the modeled distribution of states before the current measurement; the posterior p(s|m) expresses the updated distribution after that measurement. The normalizing term p(m) makes posterior probabilities sum or integrate to one. The definitions matter because an experimentally manipulated cue may alter a prior, a likelihood, or a decision rule, depending on what is learned and what the observer is asked to report. Using the same label for all these changes prevents meaningful model comparison Knill & Pouget (2004) Kersten et al. (2004).
A probability distribution contains more information than its most likely value. Its width, asymmetry, and potential multiple peaks describe uncertainty over possible perceptual causes. A point estimate is produced only after an additional rule specifies the chosen summary: a maximum a posteriori estimate, posterior mean, median, or another action tailored to the cost of different errors. These choices can diverge when distributions are skewed or multimodal. In a discrimination experiment, the response mapping and asymmetric costs may also shift the criterion independently of the posterior itself. Bayesian optimality is thus conditional on a likelihood, prior, utility function, and specified information available to the observer; an observed mean bias alone cannot reveal all four ingredients Stocker & Simoncelli (2006) Haefner et al. (2016).
A widely used simplification treats an observation and prior as Gaussian distributions around a continuous feature. In that case the posterior estimate lies between the observed value and the prior mean, weighted by their relative precisions, with precision defined as inverse variance. When measurements become less reliable, the prior generally receives greater weight. This is a mathematically explicit prediction and a useful baseline, especially in controlled cue-combination experiments. Real perception can require non-Gaussian distributions, correlated cue noise, outliers, multimodal hypotheses, and uncertainty about whether cues share a cause. An ideal-observer benchmark should therefore be treated as a falsifiable family of specified models and not a universal declaration that human performance must be optimal Ernst & Banks (2002) Stocker & Simoncelli (2006).
Generative Models and the Inverse Problem
A generative model formalizes how hidden environmental causes produce expected sensory evidence. It may contain variables for object geometry, illumination, motion, cue reliability, and causal structure, linked through conditional dependencies. Perception reverses this generative direction: it uses measurements to infer latent variables. The model's usefulness depends on whether its latent variables correspond to psychologically meaningful causes and whether the proposed statistical relations approximate the participant's relevant environment. A model fitted to laboratory displays can make elegant predictions while generalizing poorly to moving observers, rich scenes, or unfamiliar cultures of visual experience. The scientific task is to identify which assumptions the data support, which are practical simplifications, and which remain open Kersten et al. (2004) Fiser et al. (2010).
Hierarchical generative models organize hidden causes at multiple scales. A higher-level variable could represent a global scene property, while lower-level variables represent local features whose predicted activity depends on that scene. This arrangement offers a formal explanation for context sensitivity: the same local signal can be interpreted differently under different global hypotheses. Predictive coding is one possible message-passing implementation of a hierarchy, with recurrent feedback that forecasts lower-level activity and forward signals that convey deviations. The model of Rao and Ballard shows how such an architecture can reproduce some contextual receptive-field effects after natural-image learning, but the existence of a successful simulation leaves multiple biologically plausible implementations available Rao & Ballard (1999) Bastos et al. (2012).
The prior distribution itself poses a learning problem. Its appropriate form changes with experience, task, ecology, and the stability of sensory statistics. A system with a persistent slow-motion prior may produce systematic motion biases, while another task requires the observer to update transition probabilities rapidly. At a generative level, learning can be modeled as changing parameters of likelihoods or priors; at a neural level, it may involve synaptic modification, gain regulation, or the acquisition of structured representations. The challenge is to measure learning separately from transient attention or strategic response adjustment. Experimental manipulations that change environmental statistics and assess transfer to novel tests are particularly informative Stocker & Simoncelli (2006) Berkes et al. (2011) Nassar et al. (2010).
Uncertainty, Precision, and Decision Rules
Precision is the reliability assigned to a source of evidence, commonly expressed as inverse variance in a Gaussian approximation. A Bayesian observer weights more precise measurements more heavily, but the psychological meaning of precision depends on how reliability is measured. Physical contrast, stimulus duration, or sensory blur may influence uncertainty, yet their relation to internal variability must be determined empirically. Confusing the experimenter's manipulation with the observer's inferred precision can make a model seem accurate by definition. Psychometric slopes, response variability, confidence, and cue-combination weights provide partially independent estimates of the relevant uncertainty. Their convergence is stronger evidence than a post hoc fit to the average choice Ernst & Banks (2002) Alais & Burr (2004).
The normative model also specifies how actions follow posterior beliefs. Under squared-error loss, a posterior mean minimizes expected loss; under an exact-state classification rule, the most probable category may be favored; under unequal stakes, a different response criterion may be required. A participant may have a well-calibrated perceptual distribution yet use a biased report strategy, especially when reward, time pressure, or instructions alter the cost of errors. Conversely, a response pattern consistent with Bayesian integration does not establish that the nervous system explicitly represents a full probability distribution. Separating estimation, choice, and motor execution is essential when interpreting behavioral evidence Haefner et al. (2016) Nassar et al. (2010).
Precision receives a further meaning in predictive-processing theories, where the gain of error signals may control how strongly mismatches drive subsequent updates. This proposed gain regulation provides a possible bridge between attentional selection and Bayesian reliability weighting. Yet experimental attention can target locations, features, tasks, and action goals, while learned expectations alter the probabilities assigned to outcomes. They are neither operationally interchangeable nor guaranteed to modulate sensory signals in the same direction. The empirical investigation therefore requires factorial manipulations of expectation and attention, measurements before and after stimulus onset, and model comparisons that keep sensory reliability and response demands explicit Summerfield & Egner (2009) de Lange et al. (2018).
Motion Perception and the Slow-Speed Prior
The motion domain offers a particularly transparent test of probabilistic estimation because local image measurements are ambiguous and their reliability varies with contrast, spatial structure, and observation duration. Under a slow-speed prior, low-contrast motion signals receive a stronger pull toward slower velocities than relatively precise high-contrast signals. Weiss, Simoncelli, and Adelson developed an estimator that explained several apparently unrelated motion illusions as consequences of combining noisy measurements with an environmental preference for slower movement. This achievement was important because it generated quantitatively structured predictions from a compact set of assumptions. The relevant conclusion is about the explanatory power of a specified observer model, not about a universally established slow-speed prior in every organism and viewing context Weiss et al. (2002).
Subsequent work asked whether the prior and the noise function could be estimated rather than chosen simply to fit the illusion. Stocker and Simoncelli used visual-speed discrimination responses to estimate both internal noise and a prior distribution. Their fitted prior had heavier tails than a Gaussian, and noise varied with speed and contrast. This result underscores a methodological point: oversimplified likelihoods can cause researchers to attribute to prior knowledge what may actually arise from stimulus-dependent noise. Joint fitting of bias and response variance gives additional constraints, while independent measurements and held-out stimulus conditions test whether the fitted quantities retain predictive value outside the original data Stocker & Simoncelli (2006).
Motion inference also illustrates the limits of a single benchmark. Real scenes contain acceleration, self-motion, multiple moving objects, occlusion, and uncertain segmentation. A prior learned from one stimulus ensemble may be inappropriate for another, and the observer's response can depend on whether a task requires judging speed, tracking position, or planning action. Models should specify the temporal scale of their predictions and whether the relevant latent variable is retinal or world-relative motion. The successful account of a laboratory illusion motivates such extensions but does not already validate them. Cross-context generalization, measurement of eye movements, and comparison with non-Bayesian filtering models supply the next decisive evidence Weiss et al. (2002) Fiser et al. (2010).
Visual–Haptic Integration and Reliability Weighting
When vision and touch estimate the same object's height, each modality contributes information with a different precision. Under independent Gaussian errors and a common underlying property, the minimum-variance combined estimate is weighted by modality precision; a more reliable visual estimate should dominate when visual noise is small, while degraded vision should increase the contribution of touch. Ernst and Banks independently measured uncertainty within each modality and used those measurements to predict performance in their combined visual–haptic task. Observed cue weights and precision resembled the maximum-likelihood prediction. The study provided influential experimental grounding for reliability-dependent integration because the benchmark depended on empirically estimated sensory variance Ernst & Banks (2002).
The interpretation must retain the conditions under which the optimal-integration calculation applies. Independent errors, consistent calibration, compatible sources, and a stable response rule are assumptions of the simple model. If sensory measurements share noise, estimates become correlated and naive inverse-variance weighting can overcount information. If the modalities disagree sufficiently, an observer may infer separate causes rather than fuse them. Training and task instructions can also influence whether participants treat a visual signal and a tactile signal as referring to the same object. The observation of near-optimal fusion in a controlled task therefore supports a conditional computation, not a claim that all multisensory perception is reliably optimal Ernst & Banks (2002) Rohe & Noppeney (2015).
A further issue is whether integration takes place within perception or at a later stage of decision and report. A combined estimate could emerge from cortical interactions, from successive likelihood comparisons, or from a response rule that averages two already formed estimates. Distinguishing these explanations requires time-resolved measurements, task transfers, neural manipulations, and tests of whether combined cues change low-level sensory representations. The visual–haptic experiments are most compelling as behavioral evidence of statistical weighting. They leave room for theories of mechanism that make stronger, independently falsifiable claims about representational format and signal pathways Haefner et al. (2016) Aitchison & Lengyel (2017).
Audiovisual Localization and the Ventriloquist Effect
The ventriloquist effect appears when visual information shifts the apparent position of a sound. Alais and Burr measured audiovisual localization under changes in the reliability of visual localization. A clear visual signal pulled auditory location strongly toward the visual location, but substantial visual blur could reverse the dominance, allowing sound to shift the visual estimate. At intermediate reliability, responses fell between the unimodal locations, and combined localization could be more precise than either single-modality judgment. This pattern was explained by a near-optimal bimodal integration model and made reliability manipulation a central test of Bayesian-style cue combination Alais & Burr (2004).
For scientific inference, the decisive observation is the systematic relationship between independently characterized sensory reliability and the direction and magnitude of cross-modal influence. Merely showing that vision captures sound would be compatible with several simpler explanations, such as an inflexible visual preference or learned response habit. Reversing dominance by reducing visual precision makes those alternatives less plausible. However, the experimental result does not tell us that the observer has explicitly computed a posterior distribution. A fixed weighted-average heuristic with dynamically adjusted weights may reproduce some results. Distinct trialwise predictions, measurements of response variability, and conditions in which weights should be nonlinear help separate the candidate mechanisms Alais & Burr (2004) Fiser et al. (2010).
A common-cause assumption is fundamental. When an observer believes that light and sound arise from the same event, the signals can be integrated; when the discrepancy is too large or the timing is inconsistent, segregation becomes plausible. An experiment that simply measures the final cross-modal bias can therefore confuse cue weighting with causal-binding decisions. Later Bayesian causal-inference models explicitly represented uncertainty over whether signals came from one source or two. The audiovisual literature thus reveals a progression from reliability-weighted fusion toward hierarchical models in which the decision to fuse is itself inferred Rohe & Noppeney (2015).
Multisensory Causal Inference
Causal inference enters multisensory perception when sensory signals may correspond to one event or to independent events. A computational model can assign probability to a common-cause hypothesis and to separate-cause hypotheses, evaluate the likelihood of each under observed disparity and reliability, and predict either a selected percept or an average weighted across causal interpretations. The result depends on the prior probability of common origin, the discrepancy between cues, and the uncertainty attached to each cue. Importantly, a single combined estimate is not always the appropriate target. Rational segregation can preserve separate auditory and visual locations when evidence for common origin is weak Rohe & Noppeney (2015) Alais & Burr (2004).
Rohe and Noppeney studied audiovisual spatial localization using psychophysics, Bayesian model comparison, functional imaging, and multivariate analyses. They reported a hierarchy in which sensory regions predominantly reflected modality-specific estimates, posterior intraparietal regions showed signatures consistent with forced fusion, and anterior intraparietal representations corresponded more closely to causal-inference estimates. This multimethod pattern strengthens the case that uncertainty about causes matters to neural multisensory processing. It still leaves open how trial-by-trial probabilities are represented and whether the inferred hierarchy generalizes across other stimuli, movements, ages, and sensory modalities Rohe & Noppeney (2015).
Model selection becomes difficult when candidate causal models are permitted flexible priors and noise estimates. Investigators need to compare forced fusion, forced segregation, model selection, model averaging, and heuristic disparity rules on the same observations. Parameters should be estimated with uncertainty and tested in conditions designed to separate predictions, not merely optimized against a shared outcome. Conflicts between temporal and spatial cues, changing base rates of common events, and unexpected combinations can be valuable probes. Where two models remain observationally equivalent under the tested task, the conclusion should identify a supported computational equivalence class rather than single out an untested biological algorithm Aitchison & Lengyel (2017) Rohe & Noppeney (2015).
Sensorimotor Learning and Adaptive Priors
Sensory evidence is also combined with expectations during goal-directed action. Körding and Wolpert investigated sensorimotor learning in a task where uncertainty about a target and experience-dependent target statistics could influence performance. Their 2004 experiments supported behavior consistent with integrating prior information and sensory feedback in a Bayesian manner. The importance of this work is its treatment of prior knowledge as something acquired from structured experience rather than a fixed observer assumption. It connects perceptual estimation to learning and action while retaining a measurable task and explicit assumptions about uncertainty Körding & Wolpert (2004).
Adaptive priors are especially significant in volatile environments. The statistical regularity that was useful yesterday may become misleading after a sudden environmental change. A learning system must estimate not only current hidden states but also whether the generating process has changed. Nassar and colleagues developed an approximately Bayesian delta-rule account of belief updating in a changing environment, demonstrating how learning rates can vary in relation to uncertainty and environmental change. Such models provide a bridge between perceptual inference and adaptive learning, though their decision tasks should not automatically be treated as direct evidence for a single sensory cortical mechanism Nassar et al. (2010).
The next research question asks which components of adaptation are perceptual, which are decisional, and which involve motor control. Changes in pointing or reaching could result from updated sensory likelihoods, changed beliefs about target distribution, modified action policies, or altered motor noise. Experiments can distinguish these sources by collecting separate perceptual reports, varying costs independently of target statistics, and measuring aftereffects across tasks. Model-based analysis should represent the latent learning process explicitly and test recovery or transfer after the manipulation ends. The broader lesson is that reliable Bayesian-style behavior often depends on a continual interaction among sensory uncertainty, learned environmental statistics, and the consequences of action Körding & Wolpert (2004) Nassar et al. (2010).
Hierarchical Predictive Coding as a Candidate Algorithm
Predictive coding proposes that a hierarchical system transmits predicted sensory activity from higher to lower levels and residual differences in the opposite direction. In a linear model, higher-level states predict input to a lower level, and an error population represents the mismatch between actual and predicted activity. Iterative changes can then reduce the mismatch or update the estimated hidden causes. Rao and Ballard's influential 1999 model generated receptive-field-like responses and aspects of surround modulation after training with natural images. These results demonstrate computational feasibility: a particular architecture can produce certain neural response patterns. They do not establish that each biological cortical area literally contains an anatomically segregated error and prediction population Rao & Ballard (1999).
There are several families of predictive coding, distinguished by the generative model, assumed noise, objectives, error representation, and update rule. Some are derived as approximations to probabilistic inference, some emphasize efficient coding or reconstruction loss, and some expand into broader variational formulations. A successful hierarchical model may minimize squared prediction error while a richer probabilistic model weights errors by predicted precision and handles non-Gaussian latent variables. The phrase predictive processing can thus denote an explanatory family rather than one uniquely specified algorithm. Experiments can test a particular algorithm only after its variables, temporal constraints, and predicted neural signals have been stated in operational form Friston (2005) Aitchison & Lengyel (2017).
Biological interpretations require more than observing recurrent connections, because recurrence is present in many computational schemes. A theory must specify where predictions originate, how mismatches are encoded, which synapses communicate these messages, and whether neural time courses match model dynamics. The proposed message direction should be separated from the anatomical direction of axonal pathways and from effective functional connectivity estimated through data analysis. Feedforward feature detectors with lateral normalization, top-down attention, and recurrent attractors can all create context-sensitive signals. The strongest tests compare interventions and rich response distributions across models that were specified before data inspection Bastos et al. (2012) Walsh et al. (2020).
Variational Free Energy and Its Relationship to Perception
Variational formulations begin from a generative probability model and introduce an approximating distribution over hidden causes. The approximation is optimized by minimizing a quantity conventionally called variational free energy, which bounds the negative log model evidence under defined assumptions. In perception, this objective can yield update rules that reduce discrepancy between predicted and observed signals while accounting for uncertainty. Friston's 2005 theory and his 2010 review helped connect cortical inference, learning, attention, and action through this mathematical language. The formal relationship is significant, but a generic variational objective by itself does not specify which populations implement the approximation or how to observe the relevant quantities Friston (2005) Friston (2010).
The term prediction error has a precise meaning inside a particular generative model: it is the mismatch between a measurement or sufficient statistic and the value predicted by a hypothesized cause. Free-energy-based schemes frequently employ precision-weighted errors, making the weight assigned to a mismatch dependent on estimated variability. A change in the observed amplitude of a neural response can reflect altered predictions, altered precision, changing state estimates, or other physiological processes. Therefore, fitting a free-energy explanation to a response after the fact is weaker evidence than predicting the response to an independently manipulated reliability or expectation parameter before measurement Friston (2005) de Lange et al. (2018).
A further distinction concerns the level at which a theory is evaluated. Variational Bayesian inference can be an approximation method available to an engineered system even if the engineered implementation bears little resemblance to cortex. Conversely, neural circuitry can show predictive modulation without demonstrating the specific objective optimized by a formal model. Testing the connection requires deriving observable consequences of the proposed update process and comparing those consequences with alternative algorithms, including sampling, direct amortized recognition, and decision-theoretic rules. The mathematical framework offers a disciplined language for hypotheses, but empirical constraints determine its psychological and neurobiological reach Aitchison & Lengyel (2017) Keller & Mrsic-Flogel (2018).
Canonical Microcircuits and Feedback Pathways
A cortical implementation hypothesis often assigns different computational functions to ascending and descending pathways. Bastos and colleagues developed a proposal connecting predictive coding to canonical cortical microcircuits, including possible laminar origins and targets of signal exchange. The account helps translate computational variables into anatomical predictions about recurrent feedback, superficial and deep pyramidal populations, and the timing of error updates. Keller and Mrsic-Flogel later discussed converging and divergent evidence at the circuit level and emphasized the possibility of designing falsifying experiments. These works are theoretical and integrative contributions supported by selected physiological findings; their circuit assignments should be treated as testable mechanisms Bastos et al. (2012) Keller & Mrsic-Flogel (2018).
The necessary evidence includes stimulus-specific prediction signals that precede input, mismatch responses that follow violations, and selective perturbations producing the directional effects implied by the model. Laminar recordings can examine the depth profile of relevant activity, and causal manipulations can target feedback or local inhibitory populations. Yet laminar origin is not equivalent to message content. A descending signal might carry attention, motor context, or task policy, and an ascending transient might reflect adaptation, novelty, or an unexpected decision. Interpreting a signal as prediction error requires showing that it depends on a quantified mismatch between a model-derived expectation and actual sensory evidence Walsh et al. (2020) Keller & Mrsic-Flogel (2018).
Circuit models gain strength by predicting more than overall response suppression. For example, a specified prediction signal could change activity before the expected stimulus, modify gain according to reliability, and produce a distinct profile when the expected event is omitted. Such a combination makes it harder for an undifferentiated habituation account to explain all observations. Nevertheless, complicated models with many latent parameters can absorb nearly any result unless parameters are independently constrained. Research should prioritize perturbation effects, cross-condition generalization, and explicit model comparison over counting phenomena that a flexible narrative can retrospectively accommodate Bastos et al. (2012) Wacongne et al. (2012).
Expectation, Attention, and Sensory Gain
Expectation is an informed belief about the probability of sensory content, whereas attention is the selective allocation of processing resources or gain toward relevant information. The two often covary experimentally because a probable stimulus is also useful to prepare for, but they can have distinct effects. Summerfield and Egner highlighted this distinction in visual cognition, and de Lange, Heilbron, and Kok later synthesized evidence that expectations affect sensory processing at multiple stages. A design that changes probability while holding task relevance stable can probe expectation; a design that alters relevance independently of probability can examine attentional modulation. The distinction is central to interpreting apparently contradictory neural effects Summerfield & Egner (2009) de Lange et al. (2018).
Predictive theories sometimes anticipate that expected input evokes smaller mismatch signals because predicted content is efficiently explained away. Yet accurate representations can become more selective even while overall response amplitude decreases. Kok, Jehee, and de Lange combined functional imaging with multivariate analyses and found that expectation reduced the V1 response amplitude while improving information about the expected stimulus in measured patterns. The informational improvement was reported independently of the attentional effects examined, and it related to behavioral facilitation under relevant conditions. The result shows why univariate response reduction cannot automatically be equated with degraded perceptual representation Kok et al. (2012).
Other experiments found that an expected stimulus pattern could be represented even when the stimulus had not yet appeared. Kok, Failing, and de Lange examined pre-stimulus sensory templates in primary visual cortex, adding evidence that expectation can have content-specific anticipatory correlates. However, decoding a pattern does not prove that the pattern constitutes a Bayesian prior or an error-correcting prediction in the precise mechanistic sense. Preparatory attention and learned association remain considerations, depending on the experimental controls. A compelling account links the template's content, time course, causal source, and effect on subsequent behavior under independent likelihood manipulations Kok et al. (2014) de Lange et al. (2018).
Repetition Suppression, Surprise, and Prediction Error
Repeated stimulation often reduces sensory responses, while unexpected changes can increase them. Predictive-processing accounts interpret some of these effects as reduced error for predictable input and enhanced error when an established expectation is violated. A difficulty arises because repetition also produces physiological adaptation and changes in neural excitability. Probability, repetition history, and novelty are not the same variable. A rare event following a repeating sequence can be both statistically surprising and less adapted, producing a larger response under either explanation. Researchers must manipulate expectation independently from local repetition and stimulus identity to identify an error signal Walsh et al. (2020) de Lange et al. (2018).
Auditory mismatch responses provide an influential experimental setting. Wacongne, Changeux, and Dehaene constructed a layered spiking-network predictive-coding model designed to account for several features of mismatch negativity, including responses to unusual transitions and expected but omitted events. Their paper also reported a magnetoencephalography experiment relevant to the model's predictions. This combination of computational specification and new measurement is more informative than a verbal analogy between novelty and prediction error. Still, the model concerns a particular auditory circuit hypothesis, and its success does not determine the sole mechanism of all mismatch responses across brains, tasks, or developmental stages Wacongne et al. (2012).
Omission paradigms are especially valuable because the expected signal is absent and the experimenter can ask whether the system responds to a predicted event without immediate sensory drive from that event. Yet an omission response can also involve sequence timing, anticipatory attention, motor preparation, and network dynamics. Control conditions must match temporal regularity and engagement while varying the content of the missing event. Both spatial localization and temporal precision of the measured response matter. The broader research aim is to connect the sign and magnitude of a modeled sensory error with measurable neural activity and perceptual consequences under conditions that distinguish adaptation and expectation Wacongne et al. (2012) Walsh et al. (2020).
Neural Sampling and Probabilistic Population Codes
A Bayesian account of perception requires a representation sufficient for the relevant inference, but it does not uniquely prescribe how nervous tissue encodes a distribution. Candidate schemes include firing rates related to parameters of a probability distribution, temporal variability interpreted as samples, population activity encoding likelihoods, and recurrent dynamics that approximately solve an optimization problem. Knill and Pouget argued that uncertainty must be represented and used in neural computation for efficient perceptual behavior. Later theoretical work clarified that a brain performing Bayesian inference need not implement the distinctive residual-error communication proposed by every predictive-coding model. The representation and the inference algorithm are separate empirical questions Knill & Pouget (2004) Aitchison & Lengyel (2017).
Neural-sampling accounts propose that variations in neural activity across time or trials may encode uncertainty by visiting plausible latent states. Haefner, Berkes, and Fiser used a sampling-based computational framework to connect perceptual decisions, feedback, and correlated neuronal variability. Their model generated hypotheses about the task dependence of neural correlations and choice-related signals. The relevant scientific advance is that putatively noisy activity can have computational structure under an explicit probabilistic hypothesis. The challenge remains to distinguish samples from other sources of correlated variability and to identify which neural statistics correspond to trialwise distributions rather than a fitted model's convenient latent terms Haefner et al. (2016).
One line of developmental evidence concerns the organization of spontaneous cortical activity. Berkes and collaborators compared spontaneous and visually evoked activity in the developing visual cortex of awake ferrets, reporting increasing correspondence with development for natural-scene responses. A Bayesian internal-model framework predicted such convergence under learned environmental statistics. This result is suggestive of experience-shaped representations, but it does not imply that spontaneous activity has one unique interpretation as posterior sampling. Shared circuitry, developmental changes in connectivity, and state-dependent neural dynamics offer alternative contributors. Causal deprivation and altered-statistics experiments can sharpen the distinction between internal-model learning and general maturation Berkes et al. (2011) Fiser et al. (2010).
Perceptual Decisions, Confidence, and Behavioral Readouts
Behavioral studies most commonly observe a discrete response, a reaction time, a confidence rating, or a continuous estimate. Each observation is a transformed readout of a latent perceptual process. Accuracy can be high even when confidence is poorly calibrated; reaction time can change because of caution rather than a change in sensory evidence. Model comparison must therefore specify the response link, including possible decision boundaries, lapse rates, response noise, and reward structure. A posterior predictive check should reproduce distributions of responses across stimulus conditions, not simply a single mean curve. Model fits that omit the reporting process risk interpreting response strategy as prior knowledge Stocker & Simoncelli (2006) Haefner et al. (2016).
Two-alternative discrimination tasks can be especially informative because psychophysical kernels show how fluctuations in stimulus evidence influence choices at different moments. Haefner and colleagues analyzed how a sampling-based account could explain the temporal relationship among choice-related neural signals, feedback and sensory evidence. The theory suggests that a perceptual decision can itself reshape the sensory representation through recurrent inference. However, similar feedback effects can arise from attention, motor preparation, or decision-correlated signals unrelated to probability sampling. Recording populations while experimentally manipulating evidence, payoff, and feedback provides more discriminating tests than a single correlation between neural activity and choice Haefner et al. (2016).
Confidence potentially informs the posterior distribution because an observer may report not only a chosen state but also uncertainty about that state. Yet subjective confidence incorporates metacognitive access, scale use, motivation, and learning. A model predicting confidence should state whether it reads posterior probability of correctness, decision-variable magnitude, elapsed evidence, or a later evaluation. Comparing confidence and choice across independent reliability and prior manipulations can help separate accounts. The most meaningful psychological conclusion arises when a common fitted generative model successfully predicts both perceptual estimates and independent uncertainty judgments without freely adjusting parameters for each new measurement Knill & Pouget (2004) de Lange et al. (2018).
Contradictory Findings and the Perceptual Prediction Paradox
Research has sometimes described expected events as producing attenuated sensory activity, yet other experiments indicate perceptual facilitation, strengthened representation, or apparent attraction toward an expected feature. These findings demand separate measurements of response amplitude, decoding accuracy, report bias, and perceptual sensitivity. Press, Kok, and Yon examined the perceptual prediction paradox, the challenge of explaining how predictions can both suppress expected information and promote perceptual content. A single prediction-error slogan is insufficient to derive all possible outcomes. Models must specify whether a measured population predominantly reflects prediction units, error units, their mixture, or a decision-relevant representation Press et al. (2020) Kok et al. (2012).
A theoretical resolution is possible when the total neural response decreases while a smaller, better tuned subset carries more discriminable information. The V1 findings of Kok and colleagues exemplify this dissociation. Nevertheless, a plausible verbal resolution is weaker than a quantitative model that predicts the magnitude and direction of both effects for the same stimulus conditions. The researcher also needs to consider baseline normalization, stimulus adaptation, decoding analysis choices, and changes in the distribution of trials. The combined evidence supports a complex relationship between expectation and neural representation; it does not justify treating attenuation as a universal signature of successful prediction Kok et al. (2012) Walsh et al. (2020).
Yon, de Lange, and Press proposed that perceptual predictions may exhibit resistance to discrepant sensory evidence in some contexts, motivating further attention to stability, belief revision, and sensory constraints. This perspective generates questions about when a prior should be updated after repeated errors and whether strong expectations distort perception or mainly alter decision. A robust experiment would manipulate prior validity across blocks, independently degrade sensory reliability, and measure whether perception adapts in the direction predicted by a fitted learning model. The results should be compared with fixed-bias and dynamic-criterion alternatives. Such tests address the prediction paradox through observable dynamics rather than theory labels Yon et al. (2019) Nassar et al. (2010).
Alternative Algorithms and Model Identifiability
The same behavioral pattern can be produced by several computational procedures. A flexible heuristic may assign a higher weight to the more reliable cue without forming an explicit posterior; a recurrent attractor network may exhibit context sensitivity without minimizing a specified variational objective; and a feedforward discriminative network can approximate the input–output function of a Bayesian estimator after training. Computational equivalence at the level of choices does not imply mechanistic equivalence. Aitchison and Lengyel specifically examined the relationship between Bayesian inference and predictive coding, clarifying that Bayesian neural computation can occur without the canonical predictive-coding message-passing scheme. This distinction anchors responsible interpretation of model success Aitchison & Lengyel (2017).
Spratling demonstrated that particular versions of predictive coding and biased competition can be rendered mathematically equivalent under specified transformations. This result is important because a claim that predictive coding uniquely explains a pattern may be premature when another named theory makes the same prediction after expressing its variables in comparable form. Empirical discrimination requires more than fitting the same response profile with differently named models. It may require stimulus configurations where nonlinear interaction terms differ, selective disruption of feedback, or independently measured neural dynamics. Models should be compared on predictive accuracy, parameter economy, robustness, and genuinely distinct mechanistic assumptions Spratling (2008).
A useful evaluation sequence starts with a fully specified causal model, derives predictions for new stimuli, and estimates parameters on training data only. The analysis then tests held-out responses and uncertainty, compares plausible competitors under matched complexity constraints, and reports conditions of model failure. Cross-validation cannot by itself correct a missing alternative hypothesis, and an information criterion cannot determine a neural mechanism when distinct mechanisms have identical likelihoods for the measured data. Preregistered perturbations, independent replication, and recording across levels of the sensory hierarchy provide complementary evidence. Explicitly identifying what a task can discriminate is part of the scientific result Walsh et al. (2020) Aitchison & Lengyel (2017).
Experimental Design, Replication, and Scientific Inference
An informative experiment manipulates prior expectation independently of sensory likelihood. Base rates or transitional statistics can establish a probabilistic expectation, while stimulus noise, contrast, or blur controls the reliability of the incoming measurement. A factorial design tests the theoretically specified interaction: a more uncertain likelihood should generally allow a greater prior influence under the relevant model. The design needs to measure whether participants learned the intended probabilities and whether task instructions introduce reward or response biases. Trial randomization, counterbalancing, lapse modeling, and sensory calibration reduce ambiguity. If a model predicts a nonlinear interaction, the experiment must sample sufficient conditions to detect it rather than interpolate between two convenient points Stocker & Simoncelli (2006) Alais & Burr (2004).
Physiological studies add several layers of uncertainty. Functional MRI has limited temporal and cellular resolution; EEG and MEG measure aggregate signals with inverse-source ambiguity; single-unit and population recordings characterize local activity but may sample selectively from a large circuit. None of these methods directly exposes a latent posterior. Strong evidence triangulates temporal response profiles, representational content, functional consequences and causal perturbations. The critical review by Walsh and colleagues evaluates neurophysiological evidence and highlights where predictive-processing claims outrun available dissociations. A credible manuscript should use this type of critique to define the limits of neural inference rather than treat it as an obstacle to an otherwise settled theory Walsh et al. (2020).
Replication should focus on prediction-specific effects, not merely whether an experiment produces some group-level difference. Independent laboratories can reproduce the same noise manipulation, estimate cue reliability separately, and fit preregistered models with uncertainty intervals. They can also test whether results depend on training duration, response format, stimulus ecology, or analytic decisions. Negative or ambiguous outcomes become informative when the theory specifies what would have counted as evidence. Shared stimuli, transparent code, and accessible data make complex model comparisons auditable. Publication bias can inflate the apparent generality of striking prediction effects, so evidence synthesis should include null findings where available and avoid claiming a pooled effect without a documented systematic search and meta-analysis Press et al. (2020) Walsh et al. (2020).
Development, Learning, and Environmental Statistics
A prior is scientifically meaningful only if its origin and capacity for change can be investigated. Organisms encounter statistical regularities through movement, exposure, social experience, and repeated actions. Learning may alter expected feature frequencies, transition structures, causal associations, or beliefs about sensory reliability. Fiser and colleagues argued for treating perceptual inference and statistical learning as related problems because the generative model supplying the prior must itself be acquired. Developmental comparisons test whether the same regularities are present early in life or arise through experience. Longitudinal designs can separate maturation from exposure when conditions permit independently measured environmental statistics Fiser et al. (2010).
The recordings analyzed by Berkes and colleagues showed increased correspondence between spontaneous and natural-scene-evoked responses during development in ferrets, consistent with a maturing internal model of sensory statistics. Such observations make learning hypotheses empirically approachable but remain subject to alternative explanations involving developing circuitry and physiological state. An especially compelling future design would manipulate the experienced distribution of natural features, then test whether changes in spontaneous activity and perceptual bias both track the new statistics. The same fitted model should predict behavior and neural signatures across exposures. Without that joint constraint, similarities among natural stimuli, spontaneous patterns, and simulated priors remain suggestive but incompletely discriminating Berkes et al. (2011).
Research on volatile environments adds a second dimension: the organism must decide how quickly to revise a prior after change. Nassar and colleagues described an approximately Bayesian belief-updating rule whose effective learning rate depends on uncertainty and environmental change. A perceptual generalization of this logic would distinguish sensory noise, unexpected outcomes, and genuine changes in the source distribution. Behavioral adaptation might be gradual under persistent noise yet rapid after evidence for a change point. Whether the same neural mechanisms support such adjustment in visual, auditory, and sensorimotor systems remains an empirical question requiring task-specific manipulation and independent model validation Nassar et al. (2010) Körding & Wolpert (2004).
Ecological Validity and Naturalistic Perception
Laboratory psychophysics earns precision by reducing complexity, but natural perception is an active, temporally extended engagement with uncertain environments. Observers move their heads and eyes, encounter changing illumination, and gather multiple correlated cues. A prior measured in a narrow grating task may differ from the expectation guiding real-world object tracking. The ecological evaluation of Bayesian models therefore requires identifying the environmental statistics assumed by the model and asking whether observers were exposed to comparable distributions. Theoretical estimates of optimality are conditional on those assumed statistics, on available sensory information, and on the losses imposed by the task Kersten et al. (2004) Fiser et al. (2010).
Natural scenes also raise hierarchical and causal ambiguities that simplified visual tasks only partially capture. A local edge could belong to an occluded object or a background texture; a movement could reflect the observer's own motion or an external agent; an audiovisual event could represent one source or several. Generative models can formally represent these alternatives, but their parameter spaces grow rapidly with scene complexity. Learned approximate inference and discriminative shortcuts may become computationally attractive. The research question is whether a candidate model retains explanatory specificity when expanded to richer input, or whether unconstrained flexibility merely improves curve fitting Rao & Ballard (1999) Rohe & Noppeney (2015).
A stronger naturalistic research program combines controlled interventions with continuous measurement during everyday-like tasks. Eye tracking can capture information seeking, motion capture can quantify action, and calibrated audiovisual recordings can represent the stimuli encountered. Behavioral measures should be paired with explicit predictions about uncertainty and belief revision. Model comparisons must account for temporal dependence rather than treating every frame as an independent observation. When direct posterior measurement is impossible, independent assessments of cue reliability and held-out behavior strengthen inference. Such research connects the mathematical elegance of ideal observers to adaptive function while preserving observable criteria for model failure Körding & Wolpert (2004) Stocker & Simoncelli (2006).
Artificial Perceptual Systems as Comparative Research
Artificial perceptual systems provide an instructive comparison because their architectures can be inspected, trained, and perturbed in ways unavailable for human cortex. A probabilistic vision model may explicitly maintain a posterior over object location, while a neural network may learn a direct mapping from pixels to labels, and a predictive model may update latent states by reducing reconstruction error. All three can solve perceptual tasks, but their internal computations differ. Comparative research must therefore distinguish observed performance, represented uncertainty, error correction, and the particular training objective. This distinction is continuous with the methodological difference between a Bayesian description of human choices and proof of a specific neural implementation Aitchison & Lengyel (2017) Haefner et al. (2016).
In the scientific horizon of the Artificial Era, comparative perception can ask how nonbiological systems develop environmental priors, represent uncertainty, adapt under sensory shift, and combine signals from different sensors. The questions are empirically tractable when a system's data, architecture, and evaluation procedures are accessible. Calibration, robustness to rare events, causal binding under contradictory inputs, and transfer to unfamiliar environments can be assessed using controlled datasets and interventions. Claims about artificial psychological organization require additional definitions and evidence relevant to that construct. The mere existence of predictive computation establishes the presence of that computation, not every stronger psychological or philosophical attribute Knill & Pouget (2004) Rohe & Noppeney (2015).
The comparative advantage of engineered systems is mechanistic access. Researchers can disable a feedback pathway, scramble the prior, replace the sensory likelihood, or inspect whether a network carries explicit uncertainty. A model performing well after removal of putative prediction signals may challenge an interpretation that such signals were necessary for the task. Conversely, evidence that an algorithm depends on recurrent prediction errors demonstrates a mechanism in that engineered implementation and suggests testable analogies for biological work. The resulting program can refine definitions and experimental designs across human and artificial systems while keeping attribution proportional to the independently measured evidence Keller & Mrsic-Flogel (2018) Walsh et al. (2020).
Contemporary Evidence and the Mechanistic Debate Through 2026
More recent experimental work has begun to resolve which physiological processes generate responses to sensory violations. Furutachi and colleagues used a learned spatial prediction task in mice and combined cortical recordings with manipulations of thalamic and inhibitory circuitry. Their 2024 study identified a cooperative mechanism involving pulvinar input and local vasoactive-intestinal-peptide-expressing interneurons that disinhibit visual cortical neurons through a subset of somatostatin-expressing interneurons. Unexpected visual features were selectively amplified in layer 2/3 of V1. The authors argued that the observed activity encoded the unexpected sensory content more directly than a signed difference between prediction and input. This provides a concrete, causally probed mechanism for expectation-dependent sensory amplification while also challenging simplified assignments of every increased response to a canonical residual-error population Furutachi et al. (2024).
The importance of this result is methodological as well as theoretical. When a response is called a prediction error, researchers must specify whether it represents a signed subtraction, surprise magnitude, the identity of an unexpected stimulus, or a downstream decision-relevant signal. These alternatives need not have the same behavioral or neural consequences. Selective perturbation of pulvinar pathways and cortical interneurons helps establish that a particular circuit contributes to the phenomenon, yet the result is localized to a particular visual task, cortical region, and experimental species. Generalizing it into a universal account of multisensory cortex would require additional targeted tests. Furutachi and Hofer's 2026 review accordingly emphasizes distinguishing the information content of putative error signals and rethinking what sensory prediction responses can legitimately establish Furutachi et al. (2024) Furutachi & Hofer (2026).
At the level of naturalistic vision, Peelen, Berlot, and de Lange synthesized findings on scene and object processing in 2024. They described rapid partially parallel processing of objects and global scene information together with bidirectional influences through which scene context changes object recognition and object information refines scene interpretation. A joint probabilistic-inference framework offers one explanation of these reciprocal influences. Crucially, these data extend the explanatory question beyond local prediction-error suppression to structured relations among objects, scenes and contextual regularities. The neural and behavioral findings are compatible with reciprocal probabilistic constraints, while the exact probabilistic representation and message-passing algorithm remain subjects for specific tests rather than premises established by the review itself Peelen et al. (2024) Kersten et al. (2004).
Hodson, Mehta, and Smith's 2024 assessment evaluated predictive coding and active inference as algorithmic research programs. Its synthesis characterized the empirical support for predictive coding as relatively modest and identified useful but incomplete proposals for cortical implementation. The review also stressed that behavioral success by active-inference models does not settle all aspects of mechanism. These judgments matter because Bayesian inference, predictive coding, and active inference answer overlapping yet different questions and can be assessed at different levels. Their evidence should be organized by which observable phenomena the algorithms predict and which assumptions they require, especially when precision-weighting and learning parameters are fitted retrospectively Hodson et al. (2024) Aitchison & Lengyel (2017).
An explicit challenge appeared in Westerberg and Roelfsema's 2026 review, which argued that hierarchical interactions observed between sensory cortices are difficult to reconcile with influential suppressive-feedback interpretations of predictive coding. They emphasized enhancement of predicted representations by feedback and reductions associated with repetition-dependent feedforward adaptation. This is a substantive critical interpretation of the literature, and it must be compared with versions of predictive processing that predict precision-dependent response enhancement or place different messages in different populations. Furutachi and Hofer's 2026 review similarly examines varied definitions and inconsistent findings while proposing more precise tests of error content. The contemporary scientific position is best characterized by active empirical disagreement over implementation rather than by declaring a single circuit algorithm settled Westerberg & Roelfsema (2026) Furutachi & Hofer (2026) Walsh et al. (2020).
Friston and Parr's 2026 historical and theoretical review traces the lineage from Helmholtzian inference to predictive coding and active inference and offers a constructive mechanistic program. Its significance lies in showing why generative models continue to organize hypotheses even while experimental work discriminates competing forms of prediction response. The productive meeting point between this theoretical lineage and recent circuit studies is a demand for explicit measurements: the estimated latent cause, the expected sensory signal, the sign or informational content of the violation, the precision assigned to it, and the consequences of causal pathway intervention. These variables make predictive processing amenable to adjudication through future experiments and comparative artificial perceptual models Friston & Parr (2026) Furutachi et al. (2024) Westerberg & Roelfsema (2026).
Synthesis of Evidence and Unresolved Problems
The strongest behavioral evidence supports conditional Bayesian-style integration: observers often shift estimates toward prior expectations when sensory evidence is weak and combine modalities in relation to measured reliability. Motion perception, visual–haptic estimation, audiovisual localization, and sensorimotor learning provide complementary examples. Their convergence supports the usefulness of probabilistic observer models, particularly when parameters are constrained by separate measurements. The degree of generality varies with task, noise model, learning history, and decision rule. It would be inaccurate to convert the success of several paradigms into a universal claim that every perceptual report is optimal or that deviations from an optimum always have a normative explanation Weiss et al. (2002) Ernst & Banks (2002) Alais & Burr (2004) Körding & Wolpert (2004).
The strongest neural evidence supports the presence of expectation-sensitive, feedback-dependent and sometimes content-specific processing, while identifying a precise universal error-coding mechanism remains difficult. Hierarchical simulations, cortical response patterns, mismatch experiments, and physiological models make predictive coding scientifically productive. Critical reviews also demonstrate that similar observations can arise through adaptation, attention, gain control, recurrent competition or response policy. The decisive evidence should connect a specified generative model to independently measured uncertainty and causal perturbations of predicted signal pathways. A prediction-error explanation is most secure when it excludes well-defined alternatives in the conditions examined Rao & Ballard (1999) Bastos et al. (2012) Wacongne et al. (2012) Walsh et al. (2020).
Several unresolved problems therefore organize the next stage of research: identifying neural representations of full posterior uncertainty; establishing when common-cause inference is necessary rather than heuristic fusion; measuring the learning and revision of priors in natural scenes; separating predictive gain from attention; reconciling suppression with representation sharpening; and determining how robustly models transfer across ages, modalities, cultures, species, and artificial systems. Each problem calls for data that discriminate computational hypotheses, not merely broad theoretical compatibility. A rigorous research program should record contradictory results and specify the level of explanation reached by every experiment Press et al. (2020) Aitchison & Lengyel (2017) Rohe & Noppeney (2015).
Conclusion
Bayesian perceptual inference supplies a formal account of how sensory evidence and prior regularities can jointly constrain estimates of the world. Its enduring achievement is methodological: uncertainty becomes an explicit variable, and plausible interpretations can be compared through mathematically stated assumptions about likelihoods, prior distributions, and decision costs. Studies of motion, haptics, audiovisual localization, and sensorimotor learning have produced substantial evidence that human observers sometimes behave in accordance with these computations. Results remain conditional on the experiment, the estimated noise model, the learned environmental structure, and the observer's response objectives Stocker & Simoncelli (2006) Ernst & Banks (2002) Rohe & Noppeney (2015).
Predictive processing develops a family of candidate accounts of how such inference may be implemented through recurrent prediction and mismatch signaling. Hierarchical models and neural studies have revealed informative correspondences among expectation, sensory representation, circuit dynamics, and behavioral judgments. Their interpretation depends on distinguishing a successful generative explanation from proof of a unique algorithm or neural population code. Theoretical variants should state their objectives and observables explicitly, and experiments should compare predictive coding with appropriately specified alternatives. The evidence is most convincing where quantitative predictions survive changes in stimulus reliability and independently targeted causal interventions Rao & Ballard (1999) Friston (2005) Keller & Mrsic-Flogel (2018) Walsh et al. (2020). Recent direct circuit evidence indicates that unexpected visual input may be selectively amplified through thalamocortical disinhibition, and contemporary reviews disagree over whether conventional suppressive-feedback formulations capture observed sensory hierarchies. These findings demand mechanistic precision rather than an undifferentiated assumption about how error is encoded Furutachi et al. (2024) Westerberg & Roelfsema (2026) Furutachi & Hofer (2026).
The integrated scientific conclusion is that probabilistic inference is a powerful explanatory framework for perception, while the biological realization of predictive computation remains an active and increasingly testable research problem. Future progress depends on unifying measurement of priors, uncertainty and decisions with time-resolved neural data and models that generalize to new environments. Comparative artificial systems add new opportunities for mechanistic intervention and disciplined cross-system analysis. This monograph establishes Bayesian perceptual inference and predictive processing as a coherent research dossier defined by shared questions and distinct levels of explanation, with empirical constraints determining the standing of each proposed mechanism Knill & Pouget (2004) Aitchison & Lengyel (2017) Fiser et al. (2010).
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
- Alais, D., & Burr, D. (2004). The ventriloquist effect results from near-optimal bimodal integration. Current Biology, 14(3), 257–262. https://doi.org/10.1016/j.cub.2004.01.029
- Bastos, A. M., Usrey, W. M., Adams, R. A., Mangun, G. R., Fries, P., & Friston, K. J. (2012). Canonical microcircuits for predictive coding. Neuron, 76(4), 695–711. https://doi.org/10.1016/j.neuron.2012.10.038
- Berkes, P., Orbán, G., Lengyel, M., & Fiser, J. (2011). Spontaneous cortical activity reveals hallmarks of an optimal internal model of the environment. Science, 331(6013), 83–87. https://doi.org/10.1126/science.1195870
- 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(6870), 429–433. https://doi.org/10.1038/415429a
- Fiser, J., Berkes, P., Orbán, G., & Lengyel, M. (2010). Statistically optimal perception and learning: From behavior to neural representations. Trends in Cognitive Sciences, 14(3), 119–130. https://doi.org/10.1016/j.tics.2010.01.003
- Friston, K. (2005). A theory of cortical responses. Philosophical Transactions of the Royal Society B: Biological Sciences, 360(1456), 815–836. https://doi.org/10.1098/rstb.2005.1622
- Friston, K. (2010). The free-energy principle: A unified brain theory? Nature Reviews Neuroscience, 11(2), 127–138. https://doi.org/10.1038/nrn2787
- Friston, K., & Parr, T. (2026). The phantastic organ: Predictive processing in the brain. Frontiers in Psychology, 17, Article 1901293. https://doi.org/10.3389/fpsyg.2026.1901293
- 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
- Furutachi, S., Franklin, A. D., Aldea, A. M., Mrsic-Flogel, T. D., & Hofer, S. B. (2024). Cooperative thalamocortical circuit mechanism for sensory prediction errors. Nature, 633(8029), 398–406. https://doi.org/10.1038/s41586-024-07851-w
- Haefner, R. M., Berkes, P., & Fiser, J. (2016). Perceptual decision-making as probabilistic inference by neural sampling. Neuron, 90(3), 649–660. https://doi.org/10.1016/j.neuron.2016.03.020
- Hodson, R., Mehta, M., & Smith, R. (2024). The empirical status of predictive coding and active inference. Neuroscience & Biobehavioral Reviews, 157, Article 105473. https://doi.org/10.1016/j.neubiorev.2023.105473
- 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
- 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., Failing, M. F., & de Lange, F. P. (2014). Prior expectations evoke stimulus templates in the primary visual cortex. Journal of Cognitive Neuroscience, 26(7), 1546–1554. https://doi.org/10.1162/jocn_a_00562
- 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
- Körding, K. P., & Wolpert, D. M. (2004). Bayesian integration in sensorimotor learning. Nature, 427(6971), 244–247. https://doi.org/10.1038/nature02169
- Nassar, M. R., Wilson, R. C., Heasly, B., & Gold, J. I. (2010). An approximately Bayesian delta-rule model explains the dynamics of belief updating in a changing environment. Journal of Neuroscience, 30(37), 12366–12378. https://doi.org/10.1523/JNEUROSCI.0822-10.2010
- Peelen, M. V., Berlot, E., & de Lange, F. P. (2024). Predictive processing of scenes and objects. Nature Reviews Psychology, 3, 13–26. https://doi.org/10.1038/s44159-023-00254-0
- 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(1), 79–87. https://doi.org/10.1038/4580
- Rohe, T., & Noppeney, U. (2015). Cortical hierarchies perform Bayesian causal inference in multisensory perception. PLOS Biology, 13(2), e1002073. https://doi.org/10.1371/journal.pbio.1002073
- Spratling, M. W. (2008). Reconciling predictive coding and biased competition models of cortical function. Frontiers in Computational Neuroscience, 2, Article 4. https://doi.org/10.3389/neuro.10.004.2008
- Stocker, A. A., & Simoncelli, E. P. (2006). Noise characteristics and prior expectations in human visual speed perception. Nature Neuroscience, 9(4), 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
- Wacongne, C., Changeux, J.-P., & Dehaene, S. (2012). A neuronal model of predictive coding accounting for the mismatch negativity. Journal of Neuroscience, 32(11), 3665–3678. https://doi.org/10.1523/JNEUROSCI.5003-11.2012
- 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
- Weiss, Y., Simoncelli, E. P., & Adelson, E. H. (2002). Motion illusions as optimal percepts. Nature Neuroscience, 5(6), 598–604. https://doi.org/10.1038/nn0602-858
- Westerberg, J. A., & Roelfsema, P. R. (2026). Hierarchical interactions between sensory cortices defy predictive coding. Trends in Cognitive Sciences, 30(2), 110–123. https://doi.org/10.1016/j.tics.2025.09.018
- Yon, D., de Lange, F. P., & Press, C. (2019). The predictive brain as a stubborn scientist. Trends in Cognitive Sciences, 23(1), 6–8. https://doi.org/10.1016/j.tics.2018.10.003