Bottom-Up vs Top-Down Processing in Perception: Examples and Evidence

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

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

Everyday perception begins with sensory signals and operates in a world already structured by experience. We identify a familiar face despite changes in lighting, read partially obscured words, and interpret a sound more readily when the surrounding sentence makes sense. Two useful concepts explain different contributions to these achievements: bottom-up processing, the influence of incoming sensory information, and top-down processing, the influence of prior knowledge, expectations, context, and task goals. Their interaction lies at the center of modern research on perception. de Lange et al. (2018) Rauss and Pourtois (2013)

The central question is not which process wins. It is how the perceptual system weighs evidence from the present stimulus against information learned from earlier encounters. Strong, unambiguous sensory input can constrain interpretation tightly; uncertain or incomplete input often gives contextual information greater influence. Experimental findings support many effects of expectation and context while leaving open exactly which stage of processing is changed in any particular experiment. Summerfield and de Lange (2014) Firestone and Scholl (2016)

What Are Bottom-Up and Top-Down Processing?

Bottom-up processing refers to the contribution of stimulus-driven information as sensory systems register and analyze changes in light, sound, pressure, or other physical input. Top-down processing refers to influences from representations and control signals shaped by learning, expectations, goals, language, or the surrounding context. These are directions or sources of influence within interacting systems, rather than two independently operating mental organs. Different disciplines use the terms somewhat differently, which makes an explicit definition essential. Rauss and Pourtois (2013)

Bottom-up processing: information from the stimulus

For vision, light reflected from objects reaches the eyes and is transformed into neural signals. Contrast, spatial structure, edges, orientation, movement, and other features constrain possible interpretations. An abruptly appearing dark shape against a bright background supplies information even when the observer has no relevant expectation. More complex recognition gradually depends on relationships among extracted features. Such processing supplies the evidential constraints that prevent perception from being arbitrary. Gilbert and Li (2013)

Bottom-up is often described as a progression from simple features to complex representations. This is a useful starting model, but the biological system is recurrent: later visual areas send information back to earlier ones, and local connections also shape responses. Thus even early sensory activity cannot automatically be classified as a pure and isolated bottom-up stage. Gilbert and Li (2013) Rauss and Pourtois (2013)

Top-down processing: information from experience and context

Suppose a dim shape appears on a kitchen counter. Knowledge that kitchens contain mugs, bowls, and utensils changes which interpretations are likely to be considered. Familiarity with a person helps identify their face from a poor photograph. A partially masked word becomes easier to infer when the sentence strongly limits which word makes sense. These examples illustrate contextual constraint; each still depends on physical sensory evidence. Palmer (1975) de Lange et al. (2018)

Top-down influences need not be conscious intentions. Statistical regularities acquired through experience can establish expectations without deliberate reflection. An observer can anticipate a likely orientation, object category, or sound even while attending to a different task. Conversely, voluntary attention is one possible source of top-down influence, but attention and expectation differ in their operational definitions. Summerfield and de Lange (2014)

Bottom-Up vs Top-Down Processing: The Core Differences

The cleanest comparison asks four questions: Where does the information originate? What does the signal constrain? Under what conditions is it influential? How can an experiment separate its effect from other mechanisms? Bottom-up evidence is based on the incoming signal; top-down information draws on learned regularities, preceding context, and current goals. Their relative contribution changes with stimulus clarity, reliability, ambiguity, and task demands. de Lange et al. (2018)

If a clearly articulated spoken word is heard in a quiet room, the acoustic signal often specifies the word with little uncertainty. If some phonemes are masked by noise, linguistic context may narrow the alternatives. Likewise, a high-contrast outline strongly constrains the identity of a familiar object, while poor lighting invites greater use of scene information. These patterns are conditional tendencies, not rules declaring that clear signals eliminate feedback or that uncertain signals always generate accurate perception. Warren (1970) Summerfield and de Lange (2014)

Another distinction concerns the level of explanation. “Bottom-up” and “top-down” describe functional or directional contributions; “feedforward” and “feedback” are often used for patterns of communication among neural populations. They frequently correspond, but anatomy alone does not establish what a signal means psychologically. Similarly, “sensory-driven” and “automatic,” or “top-down” and “conscious,” should not be treated as synonyms. Rauss and Pourtois (2013)

Everyday Examples of Bottom-Up and Top-Down Processing

Recognizing an object in a familiar scene

Imagine seeing a small elongated shape beside a cutting board. The visual system responds to its boundaries, contrast, orientation, and surface properties. These are sensory constraints. The kitchen context makes a utensil more plausible than an equally shaped unfamiliar object. A strong contextual expectation may speed recognition when the object fits the scene; it may also lead to a mistake when the target is unexpected. Palmer’s controlled experiments on scene context provide direct historical evidence for this general phenomenon. Palmer (1975)

Reading a partially covered word

When the letters of a word are degraded, visible strokes constrain the possible letters, while the surrounding syntax and meaning constrain which candidate is likely. Reading a whole paragraph therefore makes use of local visual features and acquired linguistic structure. The demonstration is strongest when a reader is shown exactly the same ambiguous visual input in different sentence contexts. Reports of effortless comprehension by themselves cannot establish which processing stage changed. Lupyan and Clark (2015) Firestone and Scholl (2016)

Understanding speech in background noise

In conversation, acoustic timing, frequencies, and phonetic cues arrive from the speaker. When noise interrupts the signal, knowledge of language and the sentence may support an apparently continuous utterance. This is an experimentally documented form of perceptual restoration, although the listener’s response may integrate perception, interpretation, and memory. Warren (1970)

Seeing a surprising object

An unexpected object may initially generate a strong mismatch with the scene. Observers can fail to identify it, identify it slowly, or interpret it in the category most consistent with their expectations. Classic incongruent-stimulus experiments demonstrated how expectations shape reports and recognition; the particular mechanisms inferred from those observations require careful interpretation. Bruner and Postman (1949)

Hearing a syllable while watching lips

Watching visible mouth movements can change a person’s report of a simultaneous spoken syllable. The McGurk effect demonstrates that perceptual judgments about speech depend on information from more than one sensory modality. It is especially useful as a reminder that a simple division into “bottom-up sound” and “top-down knowledge” is incomplete: visual mouth movements are themselves current sensory evidence. McGurk and MacDonald (1976)

How Bottom-Up and Top-Down Information Interact

A workable psychological account describes perception as an evolving estimate of what likely produced the available evidence. Incoming signals restrict the range of plausible causes, and prior information adjusts their probabilities. Recurrent processing allows intermediate interpretations to influence how subsequent features are analyzed. The relative weighting depends partly on uncertainty: unreliable signals may benefit from prior constraints, whereas strong contradictory signals can force an interpretation to change. de Lange et al. (2018) Gilbert and Li (2013)

This account explains both the extraordinary stability and the occasional failure of perception. An expected object can be recognized efficiently under occlusion, yet an unexpected object can be overlooked or misclassified. Context can reduce uncertainty, but mistaken context introduces bias. Crucially, an error does not imply that a person simply chose what to see. It can arise at sensory, interpretive, or decision stages, which are distinct empirical possibilities. Firestone and Scholl (2016) Summerfield and de Lange (2014)

Classic Experiments: What They Demonstrate

Bruner and Postman: incongruent playing cards

In their influential 1949 work, Jerome Bruner and Leo Postman studied responses to playing cards whose color and suit arrangements violated ordinary expectations. The work showed that learned regularities can affect identification of unusual visual stimuli. Its lasting contribution is experimental attention to incongruity; historical observations alone cannot discriminate every alternative explanation involving recognition decisions, response tendencies, and the stimulus information available on each trial. Bruner and Postman (1949)

Palmer: context changes object identification

Stephen Palmer compared identification of briefly shown object drawings preceded by appropriate scenes, inappropriate scenes, or no scene. Performance and the pattern of confusions differed with context: appropriate contexts generally supported better identification than inappropriate ones. The experiment is directly relevant because the object image is tested under systematically varied contextual conditions. Nevertheless, accuracy in a forced-choice identification task is not automatically a pure measurement of visual appearance. Palmer (1975)

Warren: phonemic restoration

Richard Warren replaced a speech sound within recorded sentences with an extraneous sound. In the classic paradigm, many listeners reported a complete-sounding word or sentence and misplaced the interrupting noise, while a silent gap was more readily recognized. The experiment established a striking role for linguistic context in processing incomplete speech. It does not make the strong claim that all acoustically missing information is recreated as a veridical neural representation. Warren (1970)

McGurk and MacDonald: audiovisual speech

When an auditory syllable was paired with incongruent filmed mouth movements, observers sometimes reported a third syllable. This finding demonstrated powerful multisensory interactions in speech identification. It is evidence that perception is constructed from coordinated sources of information, not proof that beliefs overrule the senses; the conflicting input in the classic study was visual stimulation, a second bottom-up stream. McGurk and MacDonald (1976)

Kok and colleagues: expectation in early visual cortex

In an fMRI study, Kok and colleagues manipulated visual expectations and analyzed activity patterns in primary visual cortex. Expected stimuli evoked lower overall response amplitude but showed more informative multivoxel representations; the reported sharpening related to behavioral benefit when the expected feature mattered for the task. The result challenges a model in which early sensory cortex remains unaffected by expectation. At the same time, neural decoding and response amplitude are indirect measures, and one experiment does not determine the unique cellular algorithm implementing the effect. Kok et al. (2012)

Models of Perception: From Features to Predictions

Feature-based and feedforward accounts

Classical hierarchical accounts emphasize how sensory features are transformed into increasingly complex descriptions of objects. They offer a principled explanation for the dependence of recognition on contrasts, contours, and combinations of local features. Their strongest lesson is that the structure of the stimulus matters. As a full account of human perception, however, a strictly one-way pipeline has difficulty representing the extensive feedback connections and context effects documented in modern neuroscience. Gilbert and Li (2013)

Bayesian inference and reliability weighting

Bayesian formulations model the combination of prior probabilities with the likelihood of sensory data. In simplified terms, the observer estimates how likely an object or event is given both the current evidence and the statistics of earlier encounters. If the evidence is noisy, the prior can have more influence; if the evidence is highly diagnostic, a reasonable prior should be overridden. This is a mathematical way of framing probabilistic inference, not a claim that people explicitly calculate probabilities while looking at the world. Summerfield and de Lange (2014) de Lange et al. (2018)

Many observed behaviors can be captured by Bayesian-style models, but fitting a probabilistic model does not establish its unique neural implementation. Several different mechanisms can yield similar choices and response times. The distinction between a successful computational description and confirmed biological processing is particularly important in science communication. Sprevak (2023) Hodson et al. (2024)

Predictive coding and prediction errors

Predictive-coding accounts propose that higher processing levels send predictions toward lower levels and that lower levels convey information about mismatches between prediction and input. Rao and Ballard proposed an influential computational implementation that reproduced some properties of visual cortical responses. The framework makes feedback conceptually central and generates hypotheses about expectation-related responses. Rao and Ballard (1999)

Current evidence supports important effects of expectation on behavior and sensory activity, but the empirical status of predictive coding as a particular neural algorithm is more limited. Hodson and colleagues’ 2024 review described modest direct empirical support and noted that some favorable observations can also be explained by feedforward models. The broad concept of interacting sensory evidence and prior knowledge therefore has a firmer evidential basis than a claim that one predictive-coding architecture has been conclusively established. Hodson et al. (2024) Rauss and Pourtois (2013)

Perceptual organization and recurrent interpretation

Perception also groups information across space and time: edges are assigned to surfaces, parts to objects, and successive sounds to meaningful streams. Grouping principles and recurring scene regularities contribute at multiple processing stages. Recurrent feedback can help resolve local ambiguity by testing a candidate interpretation against additional sensory evidence. These explanations overlap with predictive proposals, but each should be evaluated against specific tasks and data rather than inferred from a compelling everyday demonstration. Gilbert and Li (2013) de Lange et al. (2018)

Expectation, Attention, and Sensory Evidence Are Distinct

Expectation represents an estimate of what is likely to occur. Attention allocates processing according to what is relevant or selected. A stimulus can be highly expected and irrelevant, or unexpected and central to the task. Because many experimental cues alter both expectation and task priorities, studies must manipulate them separately whenever their causal roles are at issue. Summerfield and de Lange (2014)

This distinction also protects the scientific territory of the present article. The question of how scene context and prior knowledge influence perceptual identification is different from how goals, salience, and selection history guide attention toward locations or objects. The latter belongs to a dedicated attention comparison, while this article concerns the interpretation of sensory information. Both areas interact in real experiments, and separating their definitions enables more informative designs. Rauss and Pourtois (2013)

Does Top-Down Processing Change What We See or What We Report?

A pivotal scientific dispute concerns cognitive penetrability: can higher-level beliefs, desires, or conceptual knowledge directly change perceptual experience, or do they mostly change which stimuli we inspect, which category we choose, and what we report? Firestone and Scholl argued that much reported evidence for cognitive influences on vision is compatible with alternative explanations, including attention shifts, task demands, memory, and response bias. Their critique does not deny all feedback or all context sensitivity in sensory systems; it places a demanding standard on claims that cognition alters perception itself. Firestone and Scholl (2016)

Other research emphasizes direct modulation of perceptual representations and the intimate interface between language, prediction, and perception. Reviews of expectation-related neurophysiology document changes in sensory responses that are challenging for completely encapsulated accounts. There remains no single experimental result that settles how to classify every reported effect, in part because scientists disagree about the operational boundary between perception and subsequent judgment. Lupyan and Clark (2015) Gilbert and Li (2013) de Lange et al. (2018)

An accurate interpretation identifies the measured outcome: accuracy, response time, confidence, neural activity, choice criterion, or subjective report. Saying “expectation improved recognition accuracy” is generally more defensible than concluding “belief literally changed the appearance of the object” unless an experiment separates these alternatives. This distinction is a central methodological requirement for interpreting top-down effects. Firestone and Scholl (2016)

How Scientists Measure Bottom-Up and Top-Down Influences

Stimulus quality and contextual manipulation

Researchers change sensory clarity while holding likely object identity or scene context constant. Contrast reduction, brief presentation, masking, noise, occlusion, and ambiguous contours alter bottom-up reliability. Scene primes, learned probabilities, predictive cues, and sentence frames manipulate prior or contextual information. A factorial design can cross sensory quality with expectation and test whether the context effect grows under increased uncertainty. Researchers must confirm that the cues do not inadvertently change eye movements, attention, or available sensory information. Palmer (1975) de Lange et al. (2018)

Accuracy, sensitivity, criterion, and response time

A person may say “yes, I saw it” more often because their actual ability to distinguish a target from noise improved, or because they became more willing to respond yes. Signal-detection analyses distinguish sensitivity from response criterion when task assumptions permit. Response time and confidence add information but also have multiple determinants. Without suitable control conditions, a change in correct answers cannot identify a unique perceptual mechanism. Firestone and Scholl (2016) Summerfield and de Lange (2014)

Brain imaging, electrophysiology, and causal inference

fMRI measures changes in blood-oxygen-related signals associated with neural activity; EEG and MEG provide temporally precise population-level signals; electrophysiological recording can probe neuronal responses more directly. Researchers examine whether expectations alter pre-stimulus patterns, early sensory responses, or later decision-related activity. Multivariate decoding can establish that an activity pattern contains information about a stimulus, but it does not by itself specify the subjective content of perception or prove how the information was computed. Kok et al. (2012) Gilbert and Li (2013)

The strongest designs combine converging measures with preregistered predictions, informative control conditions, and model comparisons. Reproducibility also depends on stimulus selection, participant sampling, exclusion decisions, and whether the same result appears with different tasks. A single striking illusion illustrates a phenomenon; systematic tests are required to distinguish accounts of the underlying mechanism. Hodson et al. (2024) Firestone and Scholl (2016)

When Top-Down Processing Helps and When It Misleads

Prior information can improve efficiency. It allows rapid interpretation of partially visible objects, comprehension of incomplete speech, and anticipation of likely events. In ecological settings, most regularities are useful because objects and events recur in structured environments. Context can therefore compensate for weak evidence and narrow an otherwise enormous set of candidate interpretations. de Lange et al. (2018) Warren (1970)

The same mechanism can contribute to mistaken identification when a high-probability interpretation is wrong. Unexpected objects, misleading contextual cues, and rare events may be judged using categories that ordinarily work well. Such failures depend on task conditions and should not be attributed reflexively to a particular cognitive bias or clinical condition. The normative value of an expectation depends on whether the environment actually supports it. Palmer (1975) Bruner and Postman (1949)

Experimental improvements in recognizing a briefly flashed drawing or predicting a laboratory stimulus do not show that a person can intentionally increase general intelligence, eliminate perceptual illusions, or retrain all sensory processes. Findings are specific to the stimuli, learning contingencies, task design, and performance measures that produced them. Practical claims require direct evidence in the proposed real-world setting. Hodson et al. (2024)

Development, Learning, and Individual Differences

Different observers arrive with different sensory acuity, exposure histories, languages, and learned expectations. Through development and experience they acquire increasingly informative knowledge of objects, phonemes, and typical scenes. The resulting patterns affect which hypotheses are plausible when input is incomplete. Individual variation may therefore reflect both differences in the available signal and differences in experience, rather than a general trait of being a “top-down thinker” or “bottom-up thinker.” Lupyan and Clark (2015) de Lange et al. (2018)

Experiments conducted with a narrow sample, familiar materials, and a particular language cannot automatically establish universal performance across cultures or developmental stages. Visual scenes and spoken language contain learned regularities whose prevalence can vary. Generalization requires repeated testing in appropriately sampled populations, not speculative labels assigning cognitive styles to individuals. Firestone and Scholl (2016)

Common Misunderstandings About the Two Processes

“Bottom-up is sensation and top-down is perception” is an instructional shortcut. Sensation begins with transduction, and perception includes organization and interpretation, but early sensory circuits can be affected by context, while later perceptual judgments remain constrained by sensory evidence. It is more accurate to compare sources and directions of information flow across a connected system. Gilbert and Li (2013)

“Top-down always means deliberate or conscious thinking” is another error. Probabilistic expectations may be learned and applied automatically. And “bottom-up always means correct” is equally misleading: noisy sensory input can be ambiguous or misleading before any contextual influence is considered. Both concepts describe processing contributions, not truth values. Rauss and Pourtois (2013)

“Predictive coding proves that the brain invents reality” overstates a scientific proposal. Computational predictive models are useful tools for explaining how sensory evidence and priors might be integrated. The empirical status of specific predicted neural architectures remains an active area of testing and critical comparison. Hodson et al. (2024) Sprevak (2023)

Frequently Asked Questions

What is the simplest difference between bottom-up and top-down processing?

Bottom-up processing uses information supplied by the present sensory stimulus; top-down processing uses expectations, knowledge, contextual regularities, and current goals to shape how that information is processed and interpreted. In ordinary perception the two contributions interact continuously.

Is reading an example of top-down processing?

Reading involves both. Visual information distinguishes the letters and their arrangement, while language knowledge and sentence context help resolve uncertainty. Familiar words often become easier to recognize in meaningful contexts, particularly when visual information is incomplete. Lupyan and Clark (2015)

Are visual illusions caused by top-down processing?

Some illusions involve context, expectations, or interpretation. Others can be explained substantially by relatively local stimulus relationships or sensory adaptation. An illusion is evidence of how a perceptual system behaves under specific conditions, not a universal demonstration of one top-down theory. Rauss and Pourtois (2013)

Is the McGurk effect a pure top-down effect?

No. It is a multisensory phenomenon in which current visual speech information changes responses to sound. It is valuable for showing that the senses interact, but visual input is another stream of bottom-up sensory evidence rather than necessarily a stored belief. McGurk and MacDonald (1976)

Is top-down processing the same as attention?

Attention can be a source of top-down influence, especially when a goal directs selection. Expectations concern what is probable, while attention concerns what is selected or task relevant. Experiments can dissociate them, although many everyday situations involve both. Summerfield and de Lange (2014)

Does predictive coding explain all perception?

Predictive coding is an influential family of computational and neural proposals. It explains a wide range of observations conceptually, but specific implementations face competing models and empirical tests. Scientific support for the general use of expectation is stronger than support for any single universal predictive-coding microcircuit. Hodson et al. (2024)

Can a low score on a perception task diagnose a disorder?

A laboratory score describes performance under a particular set of conditions. It is affected by sensory acuity, instructions, response strategy, fatigue, language experience, and other factors. A single task result is not a clinical diagnosis; diagnostic assessment relies on appropriate clinical criteria and broader evidence.

What the Evidence Establishes

Established findings show that context and learned expectations can affect recognition, perceptual decisions, and neural responses during sensory tasks. Classic demonstrations of scene context, phonemic restoration, and audiovisual speech provide complementary examples. Reviews of visual expectation and feedback identify multiple pathways through which prior knowledge could interact with present input. Palmer (1975) Warren (1970) McGurk and MacDonald (1976) de Lange et al. (2018)

The unresolved questions concern the precise stage at which a given effect arises, the extent to which cognition changes perceptual experience rather than reports, and which neural computations generate observed signals. Predictive-coding implementations and strong claims of cognitive penetrability remain contested or incompletely supported. A mature account treats bottom-up and top-down processing as empirically testable interacting influences, specifies the evidence for each proposed mechanism, and recognizes that perceptual judgments are shaped by both the world and the organism’s history of encountering it. Firestone and Scholl (2016) Hodson et al. (2024) Rauss and Pourtois (2013)

Related Articles

Perception: How the Mind Interprets Sensory Information establishes the wider scientific account of sensory interpretation.

Top-Down and Bottom-Up Attention: Goals, Salience, and Experience distinguishes attentional priority mechanisms from perceptual processing.

Perceptual Set: How Expectations Influence What We Perceive examines the role of prior expectations in perceptual judgment.

Predictive Processing in Perception: Theory, Evidence, and Debate reviews predictive interpretations and their disputed explanatory scope.

References

Bruner, J. S., & Postman, L. (1949). On the perception of incongruity: A paradigm. Journal of Personality, 18(2), 206–223. https://doi.org/10.1111/j.1467-6494.1949.tb01241.x

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

Firestone, C., & Scholl, B. J. (2016). Cognition does not affect perception: Evaluating the evidence for “top-down” effects. Behavioral and Brain Sciences, 39, e229. https://doi.org/10.1017/S0140525X15000965

Gilbert, C. D., & Li, W. (2013). Top-down influences on visual processing. Nature Reviews Neuroscience, 14, 350–363. https://doi.org/10.1038/nrn3476

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

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

Lupyan, G., & Clark, A. (2015). Words and the world: Predictive coding and the language-perception-cognition interface. Current Directions in Psychological Science, 24(4), 279–284. https://doi.org/10.1177/0963721415570732

McGurk, H., & MacDonald, J. (1976). Hearing lips and seeing voices. Nature, 264, 746–748. https://doi.org/10.1038/264746a0

Palmer, S. E. (1975). The effects of contextual scenes on the identification of objects. Memory & Cognition, 3(5), 519–526. https://doi.org/10.3758/BF03197524

Rao, R. P. N., & Ballard, D. H. (1999). Predictive coding in the visual cortex: A functional interpretation of some extra-classical receptive-field effects. Nature Neuroscience, 2, 79–87. https://doi.org/10.1038/4580

Rauss, K., & Pourtois, G. (2013). What is bottom-up and what is top-down in predictive coding? Frontiers in Psychology, 4, 276. https://doi.org/10.3389/fpsyg.2013.00276

Sprevak, M. (2023). An introduction to predictive processing models of perception and decision-making. Topics in Cognitive Science. https://doi.org/10.1111/tops.12704

Summerfield, C., & de Lange, F. P. (2014). Expectation in perceptual decision making: Neural and computational mechanisms. Nature Reviews Neuroscience, 15, 745–756. https://doi.org/10.1038/nrn3838

Warren, R. M. (1970). Perceptual restoration of missing speech sounds. Science, 167(3917), 392–393. https://doi.org/10.1126/science.167.3917.392