NOTE-150: Ivy & Mroczko-Wąsowicz — Framing effects in perception

Contribution

The paper introduces and defines the notion of a perceptual frame (p. 975). It argues that frames can make perceptual objects available to one perceiver and unavailable to another who receives the same input. This gives a third route, besides top-down attention and mental imagery, by which a perceiver shapes object individuation. It then uses the notion to reframe the cognitive-penetrability debate. At the level of the perceptual system over time, cognition does affect perception, because learning installs frames. At the level of an individual perceptual episode, the question is left open, because frames run bottom-up once installed (§4, §5).

Key insight

Separate the etiology of a perceptual mechanism from its operation. A frame may be installed by knowledge, goals and top-down training. Once installed, it works like a perceiver-specific “structural principle” of objecthood (p. 991) and needs no occurrent cognitive input. So “cognition shapes what we can perceive” and “perception is encapsulated at the moment of perceiving” can both be true. The authors put it as: “the cognitive origin that explains why a perceptual frame was implemented in the first place need not be the same reason why that frame operates in early sensory processing” (p. 990).

Assumptions

Premises and authorities the argument rests on:

• Structural principles of objecthood. Perceptual object individuation follows rules that fix which arrangements of parts count as objects (Green 2018, quoted p. 972): cohesion, boundedness, Gestalt laws, causal regularity. The paper stays neutral on which principles are fundamental.
• Minimal notion of perceptual object. “Perceptual objects”, “sensory individuals”, “objects for perception” and “object files” are all treated as “a singular target for perception” (pp. 971–972).
• Conservative stance for argument’s sake. Perception and cognition are assumed encapsulated (p. 971). The paper also takes no stand on whether perceptual content is rich, high-level or conceptual.
• Contrast classes. Top-down attentional perspectives are assumed to be (1) top-down directed and (2) “universally accessible to any perceiver with normally functioning capacities” (p. 974). Frames are defined against this second feature: they are not universally accessible.
• Some objects are subject-independent. Simple objects (a single dot in a lattice) are individuated by any normal perceiver (p. 974, citing Matthen 2023). So the thesis covers only a class of objects, not all of them.
• Empirical premises taken from the literature include the following. MLE cue integration weights each modality by its reliability (Ernst & Banks 2002). Prism and inversion adaptation occurs only in subjects who act in the environment (Bermejo et al. 2020 review). Amodal completion is early and perceptual: V1 involvement, initial responses about 140 ms and differential responses about 240 ms (Guttman & Kellman 2004; Murray et al. 2004). Prior object knowledge sharpens early feature detectors (Teufel et al. 2018). Experts anticipate better and use different visual information (Abernethy et al. 1994; Gredin et al. 2023).

Key results

The work is conceptual and argumentative. Its results are definitions and arguments.

• Definition (p. 975). “A perceptual frame is an adaptation of the perceptual system caused by perceptual learning and realized through bottom-up functional processes such that sensory information is organized in a subject-dependent way leading to idiosyncratic perceptual object representations.”
• §3 (pp. 975–977), frames versus imagery in expertise. Experts apply their domain search patterns outside their domain too (Ivy et al. 2023). The authors argue that on the imagery account experts would apply their images only where they fit (“You don’t see constellations where there are no stars”). They conclude that a pre-imagistic default mode, the frame, is at work. Radiologists’ performance at about 200 ms per image (Drew et al. 2013) is cited as evidence of automatic framed perception.
• §3.i, multisensory integration. A reliability-sensitive weighting principle is a “candidate” frame. Learned priors and bias corrections (Helbig & Ernst 2007; Ernst 2012) make that weighting perceiver-specific, and prism-adaptation after-effects show frames are learned. The authors conclude that frames “may be so bespoke as to yield the perception of very different perceptual objects even given the same set of sensory information” (p. 981).
• §3.ii, amodal completion (pp. 982–986). The argument has three steps. Completion represents missing parts. Without a reason to represent them they would not be represented. Therefore completion implies a reason. Against Nanay’s view that the reason is (possibly unconscious) mental imagery, the authors reread two cases as frame effects rather than imagery: aphantasic shape judgements (Jacobs et al. 2017) and red-apple priming (Kwok et al. 2019). They conclude that “where mental imagery is necessary for amodal completion, perceptual frames are necessary for imagery” (p. 985). The Teufel et al. two-tone image (Figs. 4–5) is offered as a demonstration the reader can try.
• §3.iii, anticipation. The authors follow de Vignemont (2023) in treating the anticipation of looming as amodal completion of motion. Expert-athlete anticipation data then shows that the ability to complete anticipated events depends on training, so it is framed.
• §4 (pp. 989–991), the joint. Cognition has “a direct and unmediated influence on perception insofar as it is through top-down influences that perceptual frames are encoded” (p. 990), which contradicts Firestone & Scholl’s formulation. Yet this implies neither that modularists are wrong about impenetrability nor that penetration is real. The system-level and episode-level descriptions come apart.
• §5. Frames are “something like perceiver-specific structural principles of perception” (p. 991).

Claims

id	claim	strength	support
C1	There is a class of perceiver-specific “perspectives” on a scene that cannot be reached by redirecting attention and that require training.	moderate	informal argument from cases (Bach motif, radiologist, expert athletes), plus cited expertise literature (§3)
C2	Expert perception is shaped by a pre-imagistic default mode (frame), not only by mental imagery.	weak	a single cited eye-tracking study (Ivy et al. 2023), with an inference that rests on how the imagery account would have to predict (p. 976)
C3	Frames, once installed, operate bottom-up and need no occurrent top-down influence.	weak	asserted and illustrated. §3.iii itself calls a frame “a specific top-down prediction or prior knowledge” (p. 988), and no study cited distinguishes the two
C4	Learned priors and bias corrections in multisensory integration are perceiver-specific frames that can yield different perceptual objects from the same input.	weak–moderate	cited psychophysics (Ernst & Banks 2002; Helbig & Ernst 2007; Bermejo et al. 2020). The step from “different weighting” to “different objects” is argued, not shown
C5	Every case of amodal completion involves a perceptual frame, and where imagery is needed frames are needed for imagery.	weak	a three-step argument (p. 983) whose third step is stated as a contraposition but is actually the converse, plus reinterpretation of Nanay’s two cases. See Limitations
C6	Prior object knowledge installs frames in early vision that cannot then be undone (“you cannot unsee”).	moderate	Teufel et al. (2018) data and a demonstration the reader can try (Figs. 4–5). Calling it a “frame” is the authors’ own interpretation
C7	Firestone & Scholl’s claim that perception proceeds without direct, unmediated influence from cognition is inaccurate.	informal argument	§4: cognition installs frames diachronically. This depends on reading “direct, unmediated” diachronically, which is not how the modularists meant it
C8	The system-level/episode-level distinction lets the penetrability question stay open while conceding cognitive influence on perception.	informal argument	conceptual, §4–§5. This is the paper’s main philosophical payload
C9	Framing effects count as perceptual diversity and are compatible with intersubjective objectivity and mind-independence.	assertion	one paragraph (p. 989) citing Longino 1990 and O’Callaghan 2024, not argued

Concepts

• perceptual object: minimally, “a singular target for perception”, the individual to which represented features are attributed (pp. 971–972).
• structural principles: the rules governing which arrangements count as perceptual objects and how features bind to them (p. 972).
• perceptual frame: see the definition quoted in Key results. The paper also calls it a “default perspective” that “lives in the background” and functions “as a heuristic for perceptual processing” (p. 975), and “perceiver-specific structural principles of perception” (p. 991).
• framing effect: a difference in which perceptual objects are individuated that comes from a difference in frame, not in input.
• prior knowledge (in MLE): the learned difference between veridical signal and bias. The paper stipulates that it is “an example of a perceptual frame” (p. 980).
• amodal completion: early perceptual representation of object parts that give no sensory stimulation (p. 982). The paper focuses on the top-down-influenced kind, as against completion driven by Gestalt principles.

Connections

It builds on Skrzypulec’s (2022, 2023) account of attention-constituted subject-dependent perceptual objects. It keeps the subject-dependence and drops the occurrent top-down attention. It engages Nanay (2010, 2018, 2023) on amodal completion as mental imagery, siding with Briscoe (2011) that imagery is not always required. On the perception–cognition joint it positions itself against Firestone & Scholl (2016) and with Green (2020) and Block (2023), who keep a perception/cognition border without committing to penetration. Pylyshyn is not cited, although the skim expected him. Predictive processing (Clark 2014, 2015; Hohwy 2013) supplies the vocabulary of §3.iii but not a model.

Account of mind held. The paper is representationalist about perception. Perception represents objects and attributes features to them, with object files as the working notion. It takes a broadly modular, information-processing view of the perceptual system, adopted “for the sake of argument”, on which the system is adaptive and tuned by learning. It declines to take a position on whether perceptual content is conceptual. Its distinctive position is a diachronic cognitive influence without synchronic penetration view. This is a middle position in the modularity/penetrability debate, not functionalism or interpretationism about minds as such. It fits the cognition blurb exactly: perception, perceptual learning and the architecture of mind.

Within the record, LIT-190 (Winkielman et al., cross-modal integration and the unity of consciousness) is the nearest neighbour on multisensory integration, though it concerns consciousness rather than object individuation. LIT-149 (information-processing account of representation) supplies a theory of representation this paper simply presupposes.

Bearing on the record

• Cognition tag: justified, and it should stay primary. The work is philosophy of perception and cognitive architecture throughout.
• Neuroscience tag: justifiable as secondary. The work cites V1 involvement in amodal completion, response latencies, looming detectors and cortical feedback, but only as support. It runs no neuroscience of its own.
• ML practice: none. The paper does not mention machine learning. The skim’s observation that a frame is structurally a learned inductive bias that changes what a system segments as an object is a fair analogy. It is the reader’s analogy, not the paper’s, and it carries no instruction for training or evaluating models. Nothing here belongs in the Anthology of the SOTA.
• No THEORY document in this record is supported or contradicted by it. It could seed an account of “diachronic versus synchronic cognitive influence on perception” if the record ever wants one.

Limitations

• No test separates frames from their rivals. The argument against the imagery explanation (§3) rests on a single cited study, and the authors do not show how a frame account and an attention or imagery account would predict differently in any experiment.
• The bottom-up claim is unstable. The definition says frames are “realized through bottom-up functional processes”. Yet §3.iii identifies a frame with “a specific top-down prediction or prior knowledge” (p. 988), and §3.ii cites Teufel et al.’s “top-down influences optimize early-information processing”. Predictive processing’s own story (top-down predictions at every episode) does not obviously fit “operates bottom-up”. The paper does not say what “bottom-up” means once priors are stored in low-level processing.
• The amodal-completion syllogism is invalid as written. Premise (2) is “no reason → not perceived”. Its contrapositive is “perceived → reason”. Step (3) instead states “perceived when there is a reason”, which is the converse, and then draws the correct conclusion anyway. The conclusion that completion requires a reason is fine. What does not follow is that the reason is always a frame: that is stipulated by counting learned structural preferences as frames.
• Terms are stretched by stipulation. “Where there is ‘prior knowledge,’ there is a perceptual frame at work” (p. 980), and the normal-population MLE weighting “if … counts as a perceptual frame” is one shared by nearly everyone. With these moves the concept risks covering any learned perceptual tuning. That makes the thesis easy to confirm and hard to refute.
• “Different objects” is inferred rather than shown. The multisensory evidence shows different weightings and estimates, for example of bar height. It does not show different individuation of objects.
• The anti-modularist headline trades on an equivocation. Reading Firestone & Scholl’s “direct, unmediated influence” diachronically is not how the modularists meant it. Perceptual learning is an effect those authors themselves set aside as not penetration. So the “inaccurate” verdict is weaker than it sounds.
• The section references in the text are inconsistent. The introduction promises the empirical survey “in section three”, and §3 refers back to structural principles “discussed in section one”, which is actually §2. This is cosmetic.

Open questions

• What experiment would separate a frame account from a top-down attention or unconscious-imagery account? One candidate: after-effects of trained object individuation that persist under attentional load or in aphantasics, with the individuated object, not just a weighting, as the dependent measure.
• Does the frame concept add anything beyond “perceptual learning changes low-level processing”? A criterion would be needed that marks which learned tunings are frames and which are not.
• How does “operates bottom-up” square with predictive-processing architectures in which priors are applied top-down on every episode? Is “bottom-up” here meant as “stimulus-driven given fixed parameters”?
• Does the argument extend to feature-level framing (fn. 3), and does that collapse into the cognitive-penetration-of-features debate the paper set aside?

Corrections to the seeded skim

• The skim summary (LIT-099 and NOTE-150) says frames change object individuation “without penetrating any single percept”. The paper does not claim that. It stays explicitly neutral: whether individual episodes are penetrable “remains an open question” (§4 p. 990; §5 p. 991). It also says frames are compatible with penetration (fn. 2). The accurate wording is “whether or not any single percept is penetrated”.
• The paper is stronger against modularism than the skim suggests. §4 (p. 990) calls Firestone & Scholl’s claim that perception proceeds “without any direct, unmediated influence from cognition” inaccurate, on the grounds that cognition installs frames. The skim reads the paper as only reconciling the two sides.
• The skim says §2 contrasts frames with the Necker cube, and that anyone can flip it. That contrast opens §3 (pp. 974–975). §2 sets up structural principles and attention-based subject-dependence (Skrzypulec, Ongchoco & Scholl, Carrasco).
• The paper does not hold consistently to “frames work bottom-up”. §3.iii (p. 988) sums up its evidence as the effect of “having applied a perceptual frame (a specific top-down prediction or prior knowledge)”. The skim’s clean bottom-up reading hides this tension.
• The skim implies the paper surveys empirical work on “expertise, multisensory binding, amodal completion”. The only first-party data it cites is the authors’ own eye-tracking study (Ivy et al. 2023, p. 976). Everything else is reinterpretation of published work, and it neither runs nor proposes any test that separates frames from attention or imagery.
• The keywords name predictive processing and Bayesianism, but the paper only uses them descriptively (§3.i, §3.iii). It builds no Bayesian or predictive-processing model of a frame. It glosses MLE cue combination as “Bayesian” and equates learned bias and prior with “frame” by stipulation (p. 980: “where there is ‘prior knowledge,’ there is a perceptual frame at work”).
• cognition tag: justified — it is squarely philosophy of perception and the perception–cognition joint.
• primary topic: cognition (keep). neuroscience is a fair secondary tag, since §3 draws on V1 amodal-completion latencies, looming detectors and thalamocortical feedback.