Perception

Perception is the process by which the senses give rise to experience of objects, surfaces and events. Classic theories divide over how much the mind adds to sensory input. Marr’s computational theory builds a description of the world step by step from the image; Gibson’s ecological theory holds that the information needed is already in the light and is picked up directly; Gregory’s constructivist theory treats each percept as a hypothesis built from prior knowledge (Gomes, 2000; Lobo et al., 2018; Gregory, 1997). Two disputes are active. The first is whether beliefs, desires and abilities change what people actually see, or only what they report: Firestone and Scholl find no compelling evidence that they do (Firestone & Scholl, 2016), and several claimed effects of bodily effort on perceived slope and distance failed independent tests (Hutchison & Loomis, 2006; Durgin et al., 2009). The second is whether predictive coding, a modern descendant of the constructivist view, is supported by the brain’s responses; reviews call the evidence mixed or modest (Lindsay, 2021; Hodson et al., 2024).

History

Bottom-up and direct theories

Marr proposed a bottom-up pipeline with three levels of analysis and three stages of representation: a primal sketch of edges and features, a 2.5-D sketch of visible surfaces as seen from the viewer, and a 3-D model of objects independent of viewpoint (Gomes, 2000). Gibson’s ecological approach denies that perception needs such inference or internal representation: information in the ambient optic array, the structured light reaching an observer, directly specifies affordances, the possibilities for action that the environment offers (Lobo et al., 2018). Textbooks class Gibson’s theory as direct and bottom-up (McLeod, 2022). Marr argued that Gibson seriously underrated how hard perception is, because detecting invariants in the optic array is itself a hard information-processing task (McLeod, 2022).

Constructive theories

Gregory’s constructivist view treats perceptions as predictive hypotheses, in the tradition of Helmholtz’s “unconscious inference” (Gregory, 1997). Textbooks class it as constructive and top-down: perception is a hypothesis built from prior knowledge (McLeod, 2022).

Core concepts

Object recognition

Two families of theory explain how people recognise objects from different angles. Structural-description theories, such as Biederman’s recognition-by-components, describe objects as arrangements of simple three-dimensional parts called geons; view-based theories hold that people recognise objects by comparison with stored views. The first predicts that recognition need not depend on viewpoint, under conditions Biederman and Gerhardstein specified; the second, that it does. Hayward and Tarr found viewpoint-dependent recognition even when all three of Biederman and Gerhardstein’s conditions for viewpoint invariance were met, and suggested that the features stored in views are image features, not geons (Hayward & Tarr, 1997). By 2003, the debate was described as waning, with researchers turning to how structural and view-based information combine (Hayward, 2003).

Deep convolutional networks, the image-recognition systems of machine learning, have revived the question. A 2026 preprint reports that such networks fail on new viewpoints unless trained on images from many viewpoints, and that a bias toward recognising objects by shape emerges as the diversity of viewpoints increases (Luo et al., 2026a).

Predictive coding

Predictive coding holds that the brain constantly predicts its sensory input and passes forward only the prediction error, the difference between what was predicted and what arrived. Heilbron and Chait, reviewing the auditory cortex, found strong evidence that responses are shaped by expectation and organised in a hierarchy, but no evidence for the separate populations of error neurons and prediction neurons that the theory’s strong form posits (Heilbron & Chait, 2018). Predictive processing has been criticised for lacking empirical support, and Walsh and colleagues review a recent surge of neurophysiological tests of it (Walsh et al., 2020). A review by O’Connell and Walsh judged many older studies to be weak evidence and the overall results still mixed (Lindsay, 2021). Hodson and colleagues conclude that the evidence for predictive coding is modest and that some positive results also fit feedforward models, in which information flows only from the senses upward (Hodson et al., 2024). Individual studies are mixed in a specific way: in the auditory cortex of macaques, unexpected omissions of a sound enhanced responses while unexpected repetitions did not (Shukla et al., 2026a).

Predictive coding is also linked to the dispute over top-down effects. Lupyan argues that perception is penetrable by cognition and that this follows from predictive coding; Macpherson, sympathetic to penetration, replies that Lupyan assumes predictive coding rather than arguing for it, and has not shown that the effects arise within perception rather than between separate modules (Macpherson, 2015).

Questions, controversies, and new developments

Does cognition change what we see?

This question is contested. A top-down effect, in this debate, is a change in what is perceived caused by a person’s beliefs, desires, emotions, language or abilities, rather than a change in judgment or report. Firestone and Scholl argue that the hundreds of studies claiming such effects do not provide compelling evidence for them (Firestone & Scholl, 2016). They identify six pitfalls that could produce apparent effects without any change in perception, among them effects on judgment rather than perception, the demand characteristics of the experiment (participants guessing the hypothesis and complying), differences in the visual displays, and effects of attention and memory (Collier & Lawson, 2019).

Action-specific perception

The most tested claims concern action-specific perception: the idea that people see the environment scaled by their ability to act on it, so that a hill looks steeper to someone wearing a heavy backpack, or an object looks smaller to someone who can grasp more. Firestone argues that the backpack claim is theoretically untenable even if the findings are accepted (Firestone, 2013). Critics attribute the effects to participants guessing the hypothesis and complying (Laitin et al., 2019), and independent tests have mostly supported that reading:

• When backpack wearers were given a credible alternative reason for the backpack, a cover story about measuring muscle activity, they judged slopes no steeper than people without one. The largest backpack effect came from participants who had guessed the hypothesis (Durgin et al., 2009).
• Replicating a study in which a sugary drink changed slant judgments, Durgin and colleagues concluded that the explicit requirement to wear a heavy backpack may be primarily responsible for altered judgments of hill slant (Durgin et al., 2012).
• Two experiments with three response measures failed to replicate the heavy-backpack effect on perceived distance (Hutchison & Loomis, 2006).
• Effort affected distance judgments only when instructions asked participants to include nonvisual factors, which Woods and colleagues read as a calibration of responses, not of perception (Woods et al., 2009).
• In five experiments, the effect of grasping capacity on perceived object size disappeared once size and capacity were estimated in separate tasks, and restricting the grasp changed perceived capacity but not perceived size (Collier & Lawson, 2017b). Scaling effects appeared only with unrealistically explicit instructions or when graspability was asked just before each size judgment, which the authors suggest demand characteristics could explain (Collier & Lawson, 2017a).

Proponents answer with tests designed to avoid these problems. Witt claims that four case studies of action-specific effects pass all six of Firestone and Scholl’s pitfalls (Witt, 2017). Collier and Lawson reply that Witt’s grasping case gives no support to the action-specific account: pitfalls of judgment, demand and visual differences apply to many grasping studies, though they accept Witt’s case on attention and memory (Collier & Lawson, 2019). In virtual reality, Laitin and colleagues found that a distance-on-hill effect persisted when participants received explicit feedback and vanished when both targets stood on hills, which they read as favouring a perceptual account; they describe the perceptual status of such effects as still debated (Laitin et al., 2019). Kirsch and colleagues offer a third reading: action effects on perceived size arise from shifts in spatial attention, not directly from the ability to act (Kirsch et al., 2021).

The best-supported survivor is the paddle effect, in which the size of a paddle used to hit a moving ball changes the ball’s perceived speed. It is claimed to avoid all six pitfalls, and a 2026 meta-analysis of 26 studies with 542 participants found a large pooled effect, Hedges’ g = 0.80 (Luo et al., 2026b). The same authors summarise the backpack and tool-use effects as having failed replications that support a demand-characteristics account (Luo et al., 2026b). Philbeck and Witt describe the field as polarised, with no agreement on what counts as adequate control for effects that arise after perception (Philbeck & Witt, 2015).

Replication record

Broader connections

Treating perception as statistical inference, as Bayesian models of cognition do, is itself contested: Brette rejects framing perception as inference over a predetermined model and argues that optimality is often ill defined (Brette, 2018). Action-specific perception is one of the claims of embodied cognition, which ties cognition to the body. The attentional account of action effects links this dispute to Attention. Henrich and colleagues cite cross-cultural variation in visual perception as one reason to doubt claims about human nature drawn from Western samples (Henrich et al., 2010); see Replicability and generality.

Further reading

• Firestone and Scholl (2016): the case against top-down effects on perception, and the six pitfalls.
• Philbeck and Witt (2015): the action-specific debate from both sides.
• Lobo et al. (2018): the history and philosophy of Gibson’s ecological approach.
• Heilbron and Chait (2018): what the evidence does and does not show for predictive coding in auditory cortex.