Perception

Abstract

Perception is a type of mental process, the process by which the brain organizes and interprets sensory signals into a structured representation of the environment. This article treats perception as the superordinate topic beneath which the modality-specific perceptions sit, and develops the problem that unifies them: the proximal stimulus registered by the senses underdetermines its distal cause, so perception is an inference rather than a recording. It traces psychophysics from the threshold to Stevens' power law, sets the constructivist account of Helmholtz and Gregory against Gibson's ecological alternative, reviews Gestalt organization, and shows how the Bayesian and predictive-coding frameworks recast perception as probabilistic inference that weights sensory evidence by its reliability. Three interactive demonstrations let the reader scale sensory magnitude, flip a bistable figure, and combine two sensory cues.

Keywords: perception, psychophysics, unconscious inference, bayesian inference

Perception is the set of processes that organize, identify, and interpret sensory information to build a usable representation of the environment, an achievement best understood as the information-processing problem of recovering the structure of the world from the structure of the signal it produces (Marr, 1982). It is distinguished from sensation, the transduction of physical energy into neural signals, by everything that happens after transduction: the grouping, completion, and interpretation that turn a raw signal into a perceived object at a perceived place. The distinction matters because the signal alone is never enough. A given pattern of light, sound, or pressure is consistent with indefinitely many arrangements of the world, and perception is the brain's commitment to one of them. This article is the hub for that problem in its general form; the sensory-specific cases are developed on their own pages, including Visual Perception, Auditory Perception, Depth Perception, Motion Perception, Color Perception, Face Perception, Speech Perception, Space Perception, and Time Perception. The sections below set out the inference problem, survey the modality-specific types into which perception divides, the psychophysical laws that relate stimulus to sensation, the historic contest between constructivist and ecological theories, the Gestalt principles of organization, the Bayesian synthesis that now dominates the field, and the split between perception for recognition and perception for action.

What Perception Is

The defining problem of perception is that the senses do not receive the world; they receive its effects. A three-dimensional scene projects onto the two-dimensional retina, a sound field collapses onto two eardrums, and a shape presses onto the skin at a handful of contact points. In every case the pattern that reaches the receptors, the proximal stimulus, is a degraded and ambiguous trace of the object that caused it, the distal stimulus. Recovering the distal cause from the proximal effect is mathematically an inverse problem, and inverse problems are ill-posed: many different worlds could have produced exactly the same sensory image. The retinal size of an object, for instance, is jointly determined by its physical size and its distance, so any given retinal image is consistent with a small near object or a large far one. That perception nonetheless delivers a single, stable, and usually correct answer is the fact every theory of perception must explain, and it is why perception is best understood not as reception but as the resolution of ambiguity.

Two properties of the solution recur across the modalities and organize the rest of this article. The first is perceptual constancy: perceived properties such as size, shape, lightness, and color remain roughly stable even as the proximal stimulus changes with distance, viewpoint, and illumination. Constancy is direct evidence that perception discounts the conditions of viewing to recover the object, rather than reporting the raw image. Size constancy is the direct counterpart of the size-distance ambiguity just described: the visual system holds perceived size roughly stable across distance precisely by taking the object's estimated distance into account, so that a shrinking retinal image is read as a fixed object receding rather than an object contracting in place (Gregory, 1980). The second is selectivity: only a fraction of the sensory input is elaborated into a percept at any moment, and which fraction is chosen depends on the perceiver's goals and on where Attention is deployed. Figure 1 states the inference problem schematically, and the sections that follow give the two classical answers to it and the probabilistic framework that now subsumes them.

Types of Perception

Perception is not one faculty but a family of them, divided by the sensory channel that carries the signal and by the property the system recovers from it. The Medical Subject Headings thesaurus places the descriptors below directly beneath Perception (D010465), and each is developed in full on its own page. What unites them is the inference problem set out above: in every modality the proximal signal underdetermines its distal cause, so each is a distinct instance of the same commitment to one interpretation of an ambiguous input.

Visual Perception

Visual perception is the elaboration of the retinal image into a representation of objects, surfaces, and their spatial arrangement. It is the most intensively studied modality and the source of most of the field's theory, from the constructivist and ecological accounts to the two-streams division between perception for recognition and perception for the visual control of action.

Color Perception

Color perception is the recovery of a stable surface color from light whose spectral composition varies with the illuminant. Color constancy, in which a surface keeps its apparent color as the lighting shifts, is a classic demonstration that perception discounts the conditions of viewing to recover a property of the object rather than reporting the raw signal.

Motion Perception

Motion perception is the registration of movement and change across the field, including the displacement produced by the observer's own motion. The expanding pattern of optic flow during locomotion is the invariant Gibson held to specify heading and layout directly, which made motion the paradigm case for the ecological approach.

Face Perception

Face perception is the recognition of faces and the reading of identity and expression from them. It is served by dedicated cortical machinery and has its own selective deficit, prosopagnosia, marking it as a specialized subsystem rather than a general application of object recognition.

Depth Perception

Depth perception is the recovery of the third dimension that the flat retinal image discards, and so the clearest single case of perceptual inference. It draws on binocular disparity between the two eyes and on monocular cues such as occlusion, relative size, texture gradient, and motion parallax.

Space Perception

Space perception is the apprehension of the layout of the environment and the spatial relations among its objects, integrating depth and direction into a single frame. It combines information across the senses and across successive viewpoints to build the stable spatial world in which action is planned.

Auditory Perception

Auditory perception is the organization of the sound field into distinct sources, locations, and events. The grouping and inference problems of vision recur here: the mixed pressure wave arriving at the two ears must be parsed into separate streams, an achievement known as auditory scene analysis.

Speech Perception

Speech perception is the recovery of discrete linguistic units from a continuous and highly variable acoustic signal. It is a specialized case of audition in which the same phoneme is realized differently by different speakers and contexts, so the listener must map many acoustic patterns onto one perceptual category.

Olfactory Perception

Olfactory perception is the identification of odors from the pattern of activity across a large family of receptor types. As a chemical sense it has no simple stimulus dimension analogous to wavelength or frequency, and its projections tie it closely to memory and emotion.

Gustatory Perception

Gustatory perception, or taste, is the detection of soluble chemicals in the mouth along the sweet, salty, sour, bitter, and umami dimensions. In ordinary experience it fuses with olfaction and texture into the unified percept of flavor, a clear case of multisensory integration.

Time Perception

Time perception is the estimation of duration and the ordering of events, the extension of the inference problem to the temporal dimension. It has no dedicated sense organ and is instead constructed from processes that span the modalities, which is why perceived duration stretches and compresses with attention and arousal.

Psychophysics and the Measurement of Sensation

Before perception can be explained it must be measured, and the science of relating physical stimulus to reported sensation is psychophysics, founded by Gustav Fechner in 1860. Its first achievement was the concept of the threshold. The absolute threshold is the smallest stimulus intensity that can be detected, and the difference threshold, or just-noticeable difference, is the smallest change in intensity that can be told apart from a standard. Ernst Weber had already observed that the difference threshold grows in proportion to the size of the standard: a weight must increase by a roughly constant fraction of itself to feel heavier, a regularity now called Weber's Law. Fechner integrated Weber's fraction into what became known as Fechner's law: sensation grows as the logarithm of stimulus intensity (S = k log I), so that equal ratios of intensity produce equal increments of sensation. The logarithmic law held well enough near threshold to organize a century of research, but it failed for many dimensions away from threshold. S. S. Stevens showed that direct magnitude estimation, in which observers assign numbers in proportion to how intense a stimulus seems, yields a power law rather than a logarithmic one: perceived magnitude is proportional to stimulus intensity raised to an exponent characteristic of the dimension (Stevens, 1957). The exponent is the informative quantity. It is below one for brightness and loudness, which compress a large physical range into a smaller perceptual one, and above one for electric shock, which expands it, a difference with a clear adaptive logic. The demonstration below contrasts the logarithmic and power-law accounts and lets the reader vary the Stevens exponent to see how the same physical scale is perceptually reshaped.

Constructivist Versus Ecological Theories

The oldest and most influential answer to the inverse problem is that perception supplies what the stimulus lacks by inference. Hermann von Helmholtz argued that perception is a process of unconscious inference: the visual system draws conclusions about the world from sensory data and past experience, as automatically and irresistibly as a logical deduction, and the perceiver is aware only of the conclusion, never of the premises (Helmholtz, 1867). Richard Gregory carried this constructivist tradition into modern cognitive psychology with the slogan that perceptions are hypotheses: the brain entertains a best guess about the distal scene and tests it against the incoming data, and illusions are the cases where the guess is systematically wrong (Gregory, 1980). The strongest evidence for the constructivist view is the bistable figure, a single unchanging image that the visual system perceives in two mutually exclusive ways, flipping between them without any change in the stimulus. Because the proximal stimulus is fixed while the percept alternates, the alternation can only come from the perceiver: the same data support two hypotheses, and the brain cannot commit to both at once. The demonstration below presents such a figure.

James J. Gibson rejected the whole framing. His ecological approach holds that the ambient optic array, the structured pattern of light filling the environment and transformed lawfully as the observer moves, contains enough information to specify the layout of the world without inference (Gibson, 1979). On this account the psychologist's task is not to explain how the brain enriches an impoverished image but to discover the invariants in the array, the higher-order structures such as texture gradients and optic flow that remain constant under change and directly pick out surfaces, edges, and the possibilities for action Gibson called affordances. Perception, in his phrase, is direct. The debate between constructivism and ecological realism was long treated as a stark opposition, but its modern resolution is a synthesis: the environment does carry rich structure of the kind Gibson emphasized, yet extracting and interpreting that structure under noise and ambiguity is exactly the inferential problem the constructivists described. The Bayesian account developed below formalizes both halves at once, treating the sensory information as evidence and the regularities of the world as the priors that make inference from it reliable. Table 1 sets the major theoretical approaches side by side.

Perceptual Organization

Between the raw sensory image and the recognized object lies a stage of organization: the perceptual system must decide which elements of the input belong together as a unit and which form the background against which units are seen. The Gestalt psychologists of the early twentieth century catalogued the principles by which this grouping proceeds, among them proximity, similarity, good continuation, closure, and common fate, and summarized them under the principle of Prägnanz, that the perceptual field organizes into the simplest and most stable structure the stimulus allows. A century of subsequent work has refined these principles from introspective demonstrations into quantified laws and grounded several of them in the statistics of natural scenes, but the core Gestalt insight has survived: perception is holistic before it is analytic, delivering organized wholes rather than a catalogue of independent elements (Wagemans et al., 2012). Figure-ground segregation is the most basic of these operations and the one the bistable figure above exploits, since reversing a figure and its ground is one way an image can support two incompatible organizations. Grouping also illustrates how organization can be read as inference: the principles favor the interpretations that are most probable given how surfaces and objects actually behave in the world, so that good continuation, for example, reflects the fact that contours in real scenes tend to be smooth. This convergence points toward the framework that now unifies the field.

Perception as Bayesian Inference

The modern synthesis treats perception as probabilistic inference and gives Helmholtz's unconscious inference a precise mathematical form. The perceptual system is modeled as estimating the state of the world that most probably produced the sensory data, by combining two quantities according to Bayes' rule: the likelihood, which measures how well each candidate world would account for the current sensory evidence, and the prior, which measures how probable each candidate world is before the evidence arrives. Their normalized product, the posterior, is the perceptual estimate, and the framework explains both why perception is usually veridical, because the priors encode the true statistics of the environment, and why it is occasionally fooled, because an unusual scene can be overwhelmed by a strong prior (Kersten et al., 2004). The framework's most quantitative success is in cue combination. When two cues bear on the same property, the statistically optimal estimate is a weighted average in which each cue is weighted by its reliability, the inverse of its variance, so that the more reliable cue dominates and the combined estimate is more precise than either cue alone. Marc Ernst and Martin Banks confirmed this prediction directly, showing that observers integrate visual and haptic information about size in almost exactly the reliability-weighted fashion the theory prescribes, and that experimentally degrading the visual cue shifts the weighting toward touch as predicted (Ernst & Banks, 2002). Because the visual and haptic signals come from different senses, the result also shows that perceptual inference is intrinsically multisensory, combining evidence across modalities to reduce uncertainty. David Knill and Alexandre Pouget set out the broader case that the brain represents and computes with such uncertainty as a matter of course, treating the reliability of every signal as part of the currency of neural coding (Knill & Pouget, 2004).

The Bayesian account describes what perception computes; predictive coding proposes how the brain computes it. On this scheme the cortex is a hierarchy of generative models in which each level predicts the activity of the level below and passes downward only its prediction, while the lower level returns upward only the prediction error, the part of its activity the prediction failed to explain. Rajesh Rao and Dana Ballard introduced the model to account for otherwise puzzling suppression effects in the visual cortex, where a neuron's response to a stimulus is reduced when the surrounding context makes that stimulus predictable (Rao & Ballard, 1999). The scheme naturally accommodates the pervasive top-down influences on perception, since predictions descending from higher levels shape processing at lower ones, and Charles Gilbert and Wu Li review the anatomical and physiological evidence that even primary visual cortex is continuously modulated by expectation, attention, and task (Gilbert & Li, 2013). Andy Clark has argued that predictive processing, extended from perception to action, offers a unifying principle for the mind as a whole: a brain that is fundamentally a prediction engine, perpetually matching a top-down model of its sensory causes against the incoming signal and acting to minimize the mismatch (Clark, 2013). The demonstration below realizes the cue-combination result numerically, letting the reader set the reliability of two cues and watch the optimal estimate and its precision emerge.

Perception and Action

Perception does not serve one purpose but at least two, and the visual system appears to have divided along that functional seam. Melvyn Goodale and David Milner proposed that the ventral stream running from the visual cortex into the temporal lobe supports perception in the sense of recognition and conscious report, the identification of what an object is, while the dorsal stream running into the parietal lobe supports the visual control of action, computing how to reach and grasp an object rather than merely where it is. Their account deliberately recast the dorsal stream from the earlier where pathway of spatial perception into a how pathway dedicated to the pragmatic guidance of movement (Goodale & Milner, 1992). The dissociation is supported by patients in whom brain damage spares one function while abolishing the other: a patient with ventral-stream damage may be unable to report the orientation of a slot yet post a card through it accurately, and a patient with dorsal-stream damage may describe an object perfectly yet misshape the hand in reaching for it. The two streams also differ in their susceptibility to visual illusions, with grasping sometimes resisting a size illusion that markedly distorts perceptual judgment, though the strength and interpretation of that difference remain debated. The functional split is a reminder that perception is not a single representation delivered to a central observer but a set of purpose-built processes, each constructing the description its downstream task requires. It also connects perception to the predictive-processing view, in which perception and action are two ways of reducing the same prediction error, one by revising the model and the other by changing the world to fit it (Clark, 2013).

Worked Example

The reliability-weighted combination at the heart of the Bayesian account can be made fully quantitative with the visual-haptic size task in the demonstration above. Suppose the two senses each estimate the height of a ridge. Vision reports 55 millimeters with a standard deviation of 4 millimeters, so its variance is 16; touch reports 61 millimeters with a standard deviation of 3 millimeters, so its variance is 9. The optimal estimate weights each cue by its reliability, the inverse of its variance, and normalizes. The visual reliability is one over 16, or 0.0625, and the haptic reliability is one over 9, or 0.1111, so the two weights are 0.0625 divided by 0.1736, which is 0.36, and 0.1111 divided by 0.1736, which is 0.64. Touch, being the more reliable cue, carries the larger weight. The combined estimate is therefore 0.36 multiplied by 55 plus 0.64 multiplied by 61, which is 19.8 plus 39.04, or 58.84 millimeters, pulled toward the more reliable haptic value. The precision of the combined estimate is the sum of the two reliabilities, 0.0625 plus 0.1111, or 0.1736, so its variance is one over 0.1736, which is 5.76, and its standard deviation is the square root of that, 2.4 millimeters. The crucial result is that 2.4 is smaller than either 3 or 4: combining the cues yields an estimate more precise than either sense alone could achieve, and this reduction of uncertainty is the reason perception integrates across cues and across the senses rather than deferring to any single one (Ernst & Banks, 2002). The numbers here are illustrative, but the rule, weight by reliability and the fused estimate beats its parts, is exactly what the psychophysical data show.

Discussion

Perception matters first because it is the interface on which the rest of cognition depends: every judgment, memory, and action operates on the representation perception delivers, and the systematic errors of that representation, the illusions and constancies, are not curiosities but the clearest window onto how it is built (Gregory, 1980). It matters second because it has furnished cognitive science with one of its most successful theoretical unifications. The old opposition between a constructivist psychology that saw perception as inference and an ecological psychology that saw it as the direct pickup of information has given way to a Bayesian framework that honors both, treating the world's rich structure as the source of the priors and the sensory signal as the evidence they are combined with (Kersten et al., 2004; Knill & Pouget, 2004). It matters third because that framework has descended from abstract theory to a candidate neural mechanism, predictive coding, and from there to a proposal about the organization of the mind as a whole (Rao & Ballard, 1999; Clark, 2013). The open questions are correspondingly large. It is unsettled how far the predictive-coding story is literally implemented in cortical circuitry as opposed to being one computational description among several, and it remains debated how sharply perception for recognition and perception for action are separated (Goodale & Milner, 1992). What is settled is the founding insight the whole field shares with Helmholtz: that the effortless immediacy of seeing and hearing conceals an inference of remarkable sophistication, and that the transparency of the result is precisely what makes perception worth explaining.

Commonly Confused With

Sensation

Sensation is the transduction of physical energy into neural signals by the receptors; perception is the interpretation that follows it. Apply the rule at the boundary between them: a swinging door casts a continuously changing trapezoid on the retina (sensation), yet the observer perceives a rigid rectangular door rotating in depth (perception). Wherever the two diverge is exactly where the perceptual constancies and the illusions arise, so if a phenomenon involves the stimulus being reshaped, discounted, or completed, it belongs to perception rather than to sensation.

Common Misconceptions

Perception is a faithful recording of the world, like a camera.

The sensory image is ambiguous and incomplete, and perception commits to one interpretation of it by adding assumptions, so the percept regularly goes beyond, and sometimes against, the raw input. Perceptual constancies discount viewing conditions to recover stable object properties, and illusions reveal the built-in assumptions doing the work (Gregory, 1980). The belief persists because the inference is unconscious and its result feels immediate.

The senses give the brain enough information to fix what is out there.

Recovering the distal scene from the proximal image is an ill-posed inverse problem: infinitely many world states are consistent with any given sensory image, so the data alone cannot determine a unique percept. A stable answer is possible only because the brain supplies prior knowledge of which world states are probable (Kersten et al., 2004). The impression of sufficiency comes from priors so reliable that their contribution goes unnoticed.

Perception is purely bottom-up, driven only by the incoming stimulus.

Expectation, attention, and task continuously shape perceptual processing, and this top-down influence reaches even primary sensory cortex rather than being confined to late interpretive stages (Gilbert & Li, 2013). In predictive-coding terms the brain is always matching a descending prediction against the ascending signal, so no stage of perception is ever purely stimulus-driven (Clark, 2013).

Glossary

Absolute threshold.

The smallest intensity of a stimulus that can be reliably detected, conventionally the level detected on half of the trials.

Affordance.

In Gibson's ecological theory, a possibility for action that the environment offers an organism and that the optic array specifies directly, such as a surface being walk-on-able.

Bayesian inference.

Estimation of a world state by combining a prior probability with the likelihood of the sensory evidence to yield a posterior, the framework that formalizes perception as probabilistic inference.

Bistable figure.

A single unchanging image that the visual system perceives in two mutually exclusive ways, alternating between them although the stimulus is fixed.

Cue combination.

The integration of two or more sources of information about the same property into a single estimate, optimally by weighting each cue in proportion to its reliability.

Difference threshold.

The smallest change in a stimulus that can be discriminated from a standard, also called the just-noticeable difference.

Distal stimulus.

The actual object or event in the environment that is the cause of stimulation and the target of perception.

Gestalt principles.

The rules of perceptual grouping, including proximity, similarity, good continuation, closure, and common fate, by which elements are organized into wholes.

Invariant.

In Gibson's ecological theory, a higher-order property of the optic array that stays constant as the observer moves and so directly specifies a stable feature of the world, such as a surface or an edge.

Inverse problem.

The task of recovering the distal cause from the proximal effect it produced; ill-posed because many causes are consistent with the same effect.

Optic array.

The structured pattern of light converging on a point of observation from the surrounding surfaces, whose structure Gibson held to carry the information that specifies environmental layout directly.

Perceptual constancy.

The stability of a perceived property such as size, shape, or lightness despite changes in the proximal stimulus caused by distance, viewpoint, or illumination.

Prediction error.

In predictive coding, the portion of a lower level's activity that the higher level's prediction fails to explain, and the only signal propagated upward through the cortical hierarchy.

Predictive coding.

A hierarchical scheme in which each cortical level predicts the level below and only the unexplained prediction error is passed upward, a candidate implementation of Bayesian perception.

Prior.

The probability assigned to a world state before the sensory evidence is taken into account, encoding the perceiver's knowledge of environmental regularities.

Proximal stimulus.

The pattern of energy the distal object produces at the sensory receptors, such as the retinal image, from which the distal cause must be inferred.

Psychophysics.

The quantitative study of the relation between physical stimuli and the sensations they evoke, including thresholds and magnitude scaling.

Reliability.

The inverse of the variance of a sensory estimate; the weight an optimal observer assigns a cue when combining it with others.

Stevens' power law.

The finding that perceived magnitude is proportional to stimulus intensity raised to a dimension-specific exponent, replacing Fechner's logarithmic law for magnitude estimation.

Unconscious inference.

Helmholtz's term for the automatic, non-deliberate reasoning by which the perceptual system draws conclusions about the world from sensory data and past experience.

Key Researchers

Andy Clark (b. 1957). Professor of Cognitive Philosophy at the University of Sussex; he has argued that predictive processing, in which the brain is fundamentally a prediction engine minimizing sensory error, unifies perception and action within a single account of mind. Faculty Page - ORCID - Google Scholar - Wikipedia

James J. Gibson (1904-1979). Psychologist at Cornell University and founder of the ecological approach to perception; he argued that the optic array directly specifies the layout of the environment and its affordances, so that perception need not be inferential. Wikipedia

Richard L. Gregory (1923-2010). Experimental psychologist at the University of Bristol and founding editor of the journal Perception; he developed the constructivist view that perceptions are hypotheses the brain tests against sensory data, using illusions as the key evidence. Wikipedia

Hermann von Helmholtz (1821-1894). Physicist and physiologist at the University of Berlin; his Handbuch der physiologischen Optik introduced unconscious inference, the founding idea that perception is an automatic inference from sensory data and experience. Wikipedia

David Marr (1945-1980). Vision scientist at the Massachusetts Institute of Technology; his framework of three levels of analysis and his computational theory of vision made the information-processing study of perception rigorous. Wikipedia

Frequently Asked Questions

What is the difference between sensation and perception?
Sensation is the transduction of physical energy into neural signals by the receptors, whereas perception is the organization and interpretation of those signals into a representation of objects and events in the world. The distinction is that perception adds structure the raw signal does not contain, grouping, completing, and interpreting the input to recover its distal cause (Marr, 1982).

Why is perception described as inference?
Because the proximal stimulus on the senses is consistent with many possible distal causes, an ill-posed inverse problem that the sensory data alone cannot solve. Perception reaches a single answer only by adding assumptions about which world states are probable, which Helmholtz called unconscious inference and modern theory formalizes as Bayesian estimation (Kersten et al., 2004).

What is the difference between the constructivist and ecological theories?
The constructivist view of Helmholtz and Gregory holds that the sensory image is impoverished and that perception enriches it by inference from knowledge and expectation. Gibson's ecological view holds instead that the optic array already contains information that specifies the environment directly, so that perception is the pickup of that information rather than inference from a poor image (Gibson, 1979).

What is Stevens' power law?
It is the finding that the perceived magnitude of a stimulus is proportional to its physical intensity raised to a power, an exponent that differs by sensory dimension. The exponent is below one for brightness and loudness, which compress their physical ranges, and above one for electric shock, a pattern that replaced Fechner's earlier logarithmic law for magnitude estimation (Stevens, 1957).

How does the brain combine information from different senses?
It combines cues by weighting each in proportion to its reliability, the inverse of its variance, so that the more reliable cue counts for more and the fused estimate is more precise than any single cue. Ernst and Banks showed that observers combine visual and haptic size information in almost exactly this statistically optimal way (Ernst & Banks, 2002).

Does what we expect change what we perceive?
Yes. Expectation, attention, and task modulate perceptual processing at every level, and this top-down influence reaches even primary sensory cortex rather than being limited to later interpretive stages. Predictive-coding accounts treat perception as the continuous matching of a descending prediction against the incoming signal (Gilbert & Li, 2013).

What is predictive coding?
Predictive coding is a proposal about how the cortex performs perceptual inference: each level of a processing hierarchy predicts the activity of the level below and passes down that prediction, while the lower level returns only the prediction error it could not explain. The scheme was introduced to account for context-dependent suppression in visual cortex (Rao & Ballard, 1999).

Are there separate systems for perceiving and for acting?
Evidence for two visual streams suggests so: a ventral stream supporting recognition and conscious perception and a dorsal stream supporting the visual guidance of action such as reaching and grasping. Patients with damage to one stream can lose object recognition while retaining accurate visually guided action, or the reverse (Goodale & Milner, 1992).

References