Propositional Theory vs. Analog Representation

The human mind possesses a remarkable capacity to summon absent realities: to reconstruct the layout of a childhood bedroom, rotate an intricate mechanical component within an imagined three-dimensional space, or trace the logical relationships binding an axiomatic proof. Yet beneath the effortless lucidity of conscious imagination lies one of the most foundational and passionately contested questions in the history of cognitive science: What is the formal, computational format of these internal mental representations? When an individual visualizes a setting sun or a rotating geometrical prism, does the mind operate upon an intrinsically spatial, continuous, depictive medium that preserves the physical geometry of the world? Or is this vivid sensory experience merely an introspective illusion—an experiential veneer masking a discrete, non-spatial, symbolic language of thought governed entirely by the rules of abstract propositional logic?

This fundamental inquiry gave rise to the celebrated “imagery debate,” an intellectual dispute spanning more than half a century. At its core, the debate polarizes the cognitive science community into two theoretical camps. On one side stands Propositional Theory, championed by Zenon Pylyshyn, which contends that all cognitive representations—regardless of whether they feel visual, auditory, or motoric to the experiencing subject—are fundamentally descriptive. In this view, thought is realized as discrete, language-like, truth-bearing symbols structured into predicate-argument frameworks akin to formal logic or computer code. On the other side stands the Analog (or Depictive) Framework, pioneered by Roger Shepard and Stephen Kosslyn, which asserts that the brain supports functional, quasi-pictorial representations. These depictive structures directly preserve the continuous spatial, topological, and metric properties of the represented phenomena through an internal representational medium that functions analogously to physical coordinate space.

Resolving this debate is not merely a pedantic exercise in psychological taxonomy; it strikes directly at the heart of cognitive architecture, the philosophy of mind, artificial intelligence, and the neural substrates of perception and consciousness. If propositional theory is correct, human cognition is fundamentally uniform, amodal, and computational in the classical Turing-Fodorian sense. If analog theory is correct, cognition is structurally multifaceted, retaining irreducible sensorimotor geometries that demand fundamentally different classes of neural and algorithmic computation. This treatise examines the propositional and analog paradigms in exhaustive depth, interrogating their foundational axioms, empirical paradigms, neuroscientific findings, computational implementations, and enduring philosophical implications.

1. Introduction: The Nature of Mental Representation

1.1 The Imagery Debate in Cognitive Science

The imagery debate did not emerge within an intellectual vacuum; it developed as the direct consequence of the Cognitive Revolution of the mid-twentieth century. With the rejection of radical behaviorism—which had dogmatically dismissed internal mental states as unscientific “black box” phenomena—the newly emerging cognitive science sought to build rigorous computational models of the mind. Drawing upon early computer architectures, cybernetics, and Chomskyan generative linguistics, the pioneers of cognitive science embraced the Computational Theory of Mind (CTM). Under CTM, mental operations were conceptualized as the formal manipulation of internal symbols according to algorithmic rules.

However, an acute divergence immediately arose regarding the nature of these symbols. Everyday human introspection testifies to the reality of mental imagery. When instructed to remember where one left a set of keys, an individual typically experiences a vivid, pictorial recollection of their desk or kitchen counter. The historical origin of the imagery debate crystallizes around the tension between this phenomenological reality and computational tractability. In the early 1970s, Roger Shepard and Jacqueline Metzler introduced chronometric paradigms that demonstrated that the time required to determine whether two drawn three-dimensional figures were identical was a strictly linear function of the angular disparity between them. Shortly thereafter, Stephen Kosslyn demonstrated comparable chronometric linearities in the mental scanning of spatial maps. These researchers claimed that these behavioral findings could only be explained if the mind manipulated analog, quasi-pictorial representations governed by internal metric spaces.

In 1973, Zenon Pylyshyn launched a trenchant theoretical counterattack in his paper “What the Mind’s Eye Tells the Mind’s Brain: A Critique of Mental Imagery.” Pylyshyn argued that cognitive scientists were falling victim to a seductive phenomenological fallacy. He maintained that the subjective experience of a mental “picture” has no bearing on the underlying computational architecture that actually executes the cognitive work. Pylyshyn posited that all cognitive operations are executed via a single, universal, amodal symbolic code—a language of thought. Thus began a fierce academic confrontation that transformed experimental psychology, stimulated pioneering neuroimaging paradigms, and continues to inform modern debates surrounding embodied cognition and artificial neural networks.

1.2 Epiphenomenon vs. Functional Architecture

Central to Pylyshyn’s critique is the concept of epiphenomenalism. In philosophy and cognitive science, an epiphenomenon is a secondary effect or byproduct that arises alongside a primary physical or computational process but exerts zero causal efficacy on that process. The classic analogy, originally popularized by Thomas Henry Huxley, is the steam whistle of a locomotive: the whistle produces a dramatic, perceptible auditory phenomenon, yet it plays absolutely no mechanical role in driving the wheels of the train forward.

In the context of the imagery debate, the dispute centers on whether mental images possess causal efficacy within the computational processing of the human mind, or whether they are merely the steam whistles of the cognitive engine. The analog framework posits that depictive representations are part of the functional architecture of cognition. In Kosslyn’s terminology, a functional architecture refers to the intrinsic, wired-in computational properties of the cognitive system—the hardware-level constraints and native data structures that dictate how information is stored, accessed, and transformed. If an image is part of the functional architecture, its spatial properties (such as distance, surface area, and orientation) directly constrain and determine the speed and character of cognitive computations.

Conversely, propositionalists contend that mental imagery is entirely epiphenomenal. When an individual imagines a blue square, the actual internal representation performing the cognitive computation is a collection of abstract, non-spatial propositions—such as SHAPE(SQUARE), COLOR(BLUE), and BOUNDED_BY(FOUR_EQUAL_SIDES). The introspective experience of “seeing” a blue patch in one’s mind’s eye is simply conscious readout or phenomenological noise. It plays no more functional role in the mind’s problem-solving apparatus than the glowing phosphor on a computer monitor plays in the execution of an underlying matrix multiplication inside the central processing unit.

This division holds profound implications for modularity of mind and cognitive architecture models. If imagery is epiphenomenal, the mind can be modeled as a uniform, centralized physical symbol system operating over a homogeneous syntax, fully aligning with Jerry Fodor’s classical modularity hypothesis. If imagery is functionally architectural and depictive, the mind must house specialized, domain-specific, modal subsystems whose basic operational laws differ radically from those of classical linguistic logic.

1.3 Scope and Objectives of the Comparison

To fully grasp this theoretical schism, one must synthesize philosophical, empirical, and neuroscientific viewpoints. A foundational challenge in this endeavor is establishing rigorous criteria for distinguishing a structural code from an experiential readout. The mind cannot be analyzed solely through subjective report, for introspection is notoriously vulnerable to confabulation, cognitive bias, and systematic illusions. Science requires objective, reproducible criteria: chronometric measurements, behavioral error rates, neuroimaging correlates, computational simulations, and lesions documented within clinical neuropsychology.

This comparison focuses on the foundational dichotomy that separates propositional from analog formats: discrete-abstract versus continuous-depictive. A discrete-abstract code is defined by arbitrary tokens, formal syntactic compositionality, truth-evaluability, and complete independence from any sensory modality. A continuous-depictive code is characterized by continuous spatial coordinate mappings, inherent metric topologies, topological preservation of external geometry, and intimate structural coupling with the sensorimotor apparatus.

Throughout this article, both paradigms will be evaluated on their own merits and through the critical lenses applied by their opponents. By tracking the evolution of the debate from early behavioral chronometry to contemporary functional magnetic resonance imaging (fMRI), transcranial magnetic stimulation (TMS), and deep artificial neural networks, this analysis clarifies how the human mind achieves the astonishing feat of internal representation.

2. Propositional Theory: The Descriptive Framework

2.1 Foundational Principles and Pylyshyn’s Formulation

Propositional Theory rests upon the assertion that the currency of human thought is fundamentally uniform, abstract, and language-like. Championed by philosophers such as Jerry Fodor and cognitive scientists such as Zenon Pylyshyn, this perspective is conceptually tethered to the Language of Thought (LOT) hypothesis, frequently termed Mentalese. The foundational principle of this framework is that cognitive processes consist of syntactic operations executed over symbolic structures. These structures do not physically or functionally resemble the objects they represent; rather, they denote them through established, arbitrary conventions, precisely as words in natural language denote their referents.

Pylyshyn’s fundamental formulation rejects the notion that the human brain can harbor intrinsic spatial properties inside its mental storage mechanisms. In his classic work, Pylyshyn argued that postulating internal “pictures” creates severe philosophical and computational absurdities. Foremost among these is the notorious homunculus fallacy: if the mind constructs a mental picture, there must be an internal agent—a little person or homunculus—stationed within the brain to view, inspect, and interpret that picture. This immediately triggers an infinite regress, because one must then explain how that internal homunculus processes visual information, requiring yet another homunculus inside its own mind, ad infinitum.

To avoid this conceptual trap, Pylyshyn insisted that mental representations must be fundamentally conceptual, non-perceptual, and descriptive. Thought does not occur within a “Cartesian theater” where images are projected onto an internal screen. Instead, the brain stores information as structured networks of semantic relations. When an individual engages in what feels like mental visualization, the cognitive engine is merely retrieving and executing a series of symbolic descriptions. The pictorial sensation accompanying this retrieval is dismissed as an incidental byproduct: a phenomenological illusion that carries no informational or computational utility.

2.2 Syntactic Structure, Predicate Calculus, and Abstract Semantics

To understand how Propositional Theory operates computationally, one must examine its syntactic structure, which is derived from formal logic, particularly predicate calculus. A proposition is defined as the smallest unit of knowledge that can stand as an independent, truth-evaluable assertion. It is capable of being judged true or false with respect to a designated state of affairs. In propositional systems, knowledge is represented via predicate-argument structures of the general form:

PREDICATE(ARGUMENT_1, ARGUMENT_2, ..., ARGUMENT_N)

For example, rather than storing a continuous visual depiction of a coffee cup resting upon an oak table, a propositional cognitive architecture encodes the relationship through discrete tokens such as ON(CUP, TABLE). If the scene contains additional metric and qualitative specifics, these are appended via discrete, hierarchically nested propositions:

• COLOR(CUP, WHITE)
• MATERIAL(TABLE, OAK)
• SPATIAL_RELATION(CUP, RIGHT_OF, LAPTOP)
• DISTANCE(CUP, LAPTOP, 15_CENTIMETERS)

This format is intrinsically symbolic, discrete, and compositional. The meaning of a complex propositional structure is entirely a function of the semantic values of its individual symbolic tokens and the syntactic rules used to combine them, adhering strictly to the principle of compositionality formulated by Gottlob Frege. A critical attribute of this representational format is its modality invariance. The proposition ABOVE(PICTURE_FRAME, FIREPLACE) is neither visual nor linguistic nor auditory. It is an abstract conceptual token that can be accessed with equal facility by the visual recognition system, the linguistic production apparatus, or the motor planning network. The symbolic token FIREPLACE does not smoke, possess warmth, or occupy space; it serves as an arbitrary pointer to an entry in the mind’s semantic memory network.

Furthermore, propositional representations naturally preserve categorical distinctions while discarding irrelevant metric ambiguities. In an analog picture of a dog standing beside a tree, one cannot depict the dog without specifying its exact posture, its precise distance from the trunk, the angle of illumination, and its scale relative to the bark. A propositional representation, however, excels at abstract economy: it can simply encode NEAR(DOG, TREE) without committing to arbitrary metric coordinates. This conceptual abstraction is arguably vital for higher-order reasoning, categorical generalization, and the construction of counterfactual inferences.

2.3 Explaining Behavioral Data via Tacit Knowledge

A primary challenge for Propositional Theory was explaining the robust behavioral findings generated by analog imagery researchers. For instance, in the mental scanning experiments conducted by Kosslyn and colleagues, human participants took longer to mentally “scan” across greater distances on an imagined map than across shorter distances. To the analog camp, this was clear proof that the mind was traversing an internal metric space. How could a purely propositional system, which retrieves symbolic nodes instantaneously via associative links, account for this systematically linear reaction time?

Pylyshyn responded with the tacit knowledge hypothesis. He argued that human participants in imagery experiments do not reveal the architectural constraints of their cognitive hardware; rather, they unconsciously simulate what they know would happen in a real-world physical scenario. Humans possess extensive tacit knowledge regarding the laws of physics, sensory perception, and spatial movement. They know that in physical reality, moving one’s eyes or attention across a vast physical distance takes more time than scanning a minuscule distance. When an experimenter explicitly instructs a participant to “imagine a map, focus on the tower, and then scan your eyes across to the beach,” the participant utilizes their tacit knowledge to comply with the experiment’s pragmatic context.

Under this view, the reaction-time delays reflect cognitive penetrability. A process is cognitively penetrable if it can be systematically altered, modified, or abolished by changing an individual’s explicit beliefs, goals, instructions, or expectations. Pylyshyn demonstrated that if the task instructions are altered—for example, if participants are told to simply report whether a specific feature exists on the map without being cued to simulate the visual movement of an attention point—the distance-to-latency scanning effect often evaporates completely. Thus, the variation in response latencies does not stem from an analog medium enforcing a metric velocity limit; it is generated by the participant methodically parsing their propositional database at a cadence that accurately mimics the temporal dynamics of physical observation.

3. Analog Representation: The Depictive Framework

3.1 Foundational Principles and Kosslyn’s Quasi-Pictorial Model

The Analog (or Depictive) Framework operates on an entirely distinct set of computational premises. It asserts that certain cognitive tasks are accomplished through internal representations that preserve the inherent spatial and topological metrics of the physical objects they represent. In an analog representation, the physical or functional medium itself embodies the relational dimensions of the world. Just as a physical cartographic map represents the distance between London and Paris through the proportional distance of ink on paper, an analog mental representation preserves spatial relations through an internal metric coordinate space.

The definitive computational formalization of this framework was constructed by Stephen Kosslyn in his seminal works, Image and Mind (1980) and Image and Brain (1994). Kosslyn explicitly rejected the homunculus fallacy by demonstrating that depictive representations do not require an internal observer. Instead, they function as two-dimensional or three-dimensional coordinate arrays within a specialized computational subsystem termed the visual buffer.

In Kosslyn’s architecture, the visual buffer is a functional array configured such that each representational node or “pixel” corresponds to a specific point on the surface of the represented entity, preserving neighborhood topologies and metric adjacency. The information is depictive because the medium enforces structural constraints: points that are close together in the physical object are represented by cells that are topologically proximate within the buffer. The interpretive computational processes operating over this buffer do not read text-like assertions; rather, they perform continuous transformations, spatial cropping, boundary tracing, and metric scanning directly on the coordinate topology.

Kosslyn divided mental representations into two distinct structural types:

• Deep representations: Abstract, long-term memory traces stored in either categorical propositional formats or associative metric descriptions within temporal and parietal structures.
• Surface representations: Transient, quasi-pictorial patterns actively generated within the short-term visual buffer. These are the functional entities that correspond to active mental images.

Critically, the visual buffer is not a magical canvas; it is a specialized, functionally embodied register that shares computational architecture with the human visual processing system, particularly the retinotopically organized regions of the cerebral cortex.

3.2 Empirical Paradigms Supporting Depictive Formats

The credibility of the depictive framework rests upon a vast, rigorously replicated corpus of chronometric behavioral experiments. The foundational breakthrough occurred with the classic work of Roger Shepard and Jacqueline Metzler (1971) on mental rotation. Participants were presented with pairs of perspective line drawings depicting three-dimensional, multi-armed asymmetric block configurations. Their task was to judge whether the two drawings represented identical objects or stereoisomers (mirror-image inversions). The angular disparity between the figures was parametrically varied across both the picture plane (two-dimensional rotation) and in depth (three-dimensional rotation).

The results were unequivocal: the reaction time required to reach a decision increased as a direct, strictly linear function of the angular disparity between the objects. This linear relationship held true whether the rotation occurred within the picture plane or in depth, with participants rotating the representations at a characteristic velocity of approximately 60 degrees per second. Shepard concluded that to compare the two objects, participants performed an internal, analog simulation of continuous physical rotation. The mental representation had to pass through all intermediate angular orientations (e.g., 30°, 60°, 90°) before reaching the target alignment (180°). In a purely propositional system, rotating an object computationally should involve replacing a single symbolic argument (e.g., changing ORIENTATION(0) to ORIENTATION(180)), which should execute in constant time regardless of the numerical magnitude of the angle. The discovery of continuous, angle-dependent latencies provided profound empirical support for analog computation.

Shortly thereafter, Kosslyn, Ball, and Reiser (1978) designed the canonical fictional island mental scanning paradigm. Participants meticulously memorized a map of a fictional island containing seven distinct landmarks (e.g., a hut, a tree, a well, a beach) until they could draw their locations with absolute spatial precision. The physical distances between these landmarks varied widely across the map. During testing, the physical map was removed, and participants were instructed to construct a mental image of the island, focus their mental attention on a designated starting landmark (e.g., the hut), and wait for the name of a second landmark (e.g., the beach). Upon hearing the second landmark, they were instructed to mentally “scan” across their mental image from the first point to the second, pressing a button the instant their imagined focus arrived at the destination.

The outcome was striking: the time taken to scan across the mental image was a direct, linear function of the actual physical distance between the landmarks on the original map. This linear latency function emerged despite the fact that participants knew all landmarks were situated on the same small island. Kosslyn argued that this distance-to-latency ratio could only occur if the mental image structurally preserved the metric properties of the original visual stimulus within an analog medium where greater distance functionally requires greater traversal time.

A third compelling line of behavioral evidence emerged from studies examining image size and visual acuity (Kosslyn, 1975). Participants were instructed to mentally visualize a target animal (such as a rabbit) positioned adjacent to either a colossal animal (such as an elephant) or a minuscule creature (such as a fly). While maintaining this mental image, participants were asked feature-verification questions, such as “Does the rabbit have whiskers?” Kosslyn discovered that participants were significantly faster to verify fine visual details when the target animal was imagined next to the tiny fly than when imagined next to the massive elephant. Because the elephant dominated the capacity of the active visual buffer, the imagined rabbit had to be represented at a subjective scale with minimal visual acuity, forcing the cognitive system to execute an explicit “zoom” operation before resolving fine features. This size-dependent resolution directly mirrors the optical constraints of biological vision, presenting a profound theoretical challenge to propositional models that operate on discrete semantic links devoid of resolution thresholds.

3.3 Neurocomputational and Neuroimaging Evidence

While behavioral chronometry provided powerful indirect evidence for analog representation, the advent of modern cognitive neuroscience shifted the debate into the empirical arena of neurobiology. If analog representations are real, they must be realized within neural substrates that possess the appropriate structural architecture to host depictive spatial arrays. That architecture exists: the retinotopic maps of the primary visual cortex (Brodmann Area 17, or V1) and adjacent early visual cortices (V2, V3).

Retinotopy is a neuroanatomical organization in which neighboring projection neurons on the human retina preserve their spatial topology as they map onto neighboring cortical columns in V1. If an individual looks at a bright circle, an actual geometric circle of electrical activation forms across the cortical surface of V1, distorted only by a logarithmic cortical magnification factor that devotes vastly more cortical territory to the central fovea than to the visual periphery (as demonstrated in the definitive primate autoradiography studies by Roger Tootell and colleagues).

Stephen Kosslyn, William Thompson, and their collaborators leveraged functional neuroimaging (PET and fMRI) to test whether this retinotopic architecture is recruited during pure, voluntary mental imagery in the total absence of physical visual input. In landmark experiments, Kosslyn et al. (1995, 1999) demonstrated that when human subjects close their eyes and generate high-resolution visual mental images, primary visual area V1 exhibits statistically significant, spatially localized metabolic activation. Crucially, this activation is systematically organized: visualizing an object at a large imagined visual angle activates the anterior, peripheral representations of V1, whereas visualizing the same object at a small imagined visual angle restricts activation to the posterior, foveal representations of V1.

To establish that this V1 activation is not merely another epiphenomenal neural byproduct, Kosslyn et al. (1999) deployed repetitive Transcranial Magnetic Stimulation (rTMS) to deliver transient, non-invasive disruptive magnetic pulses directly over the primary visual cortex of human subjects. The participants performed visual feature comparison tasks under two conditions: a perceptual viewing condition and an eyes-closed mental imagery condition. Disruption of Area 17 via rTMS significantly impaired performance and increased response latencies in both the perceptual task and the mental imagery task, whereas sham stimulation over control brain regions produced no such impairment. This demonstrated causal necessity: if disrupting the retinotopic computational fabric of Area 17 disrupts mental imagery processing, the depictive medium is not an epiphenomenon. It is an indispensable computational engine of the cognitive task.

Furthermore, neuropsychological lesion studies consistently reinforce this shared neural architecture. Patients who suffer localized damage to the occipital lobes resulting in visual field deficits (such as homonymous hemianopia, the loss of half the visual field) frequently display an identical, co-occurring loss in their imagined visual field. For instance, in an iconic clinical case documented by Martha Farah (1984), a patient who underwent unilateral occipital lobectomy demonstrated a dramatic reduction in the maximal horizontal extent of her mental visual images; her imagined “visual angle” was cut precisely in half, matching the exact sensory perimeter of her blinded perceptual field.

4. Key Differences Between Propositional and Analog Representations

The fundamental divergences between Propositional Theory and the Analog Framework are structural, functional, computational, and neurobiological. The following comprehensive comparison table delineates the essential dimensions separating the two theoretical models:

Dimension / Aspect	Propositional Representation	Analog Representation
Core Representational Format	Discrete, symbolic tokens organized into predicate-argument structures and semantic networks.	Continuous, quasi-pictorial coordinate arrays that preserve metric and topological spatial relations.
Semiotics and Referent Link	Arbitrary; the symbol (e.g., TREE) has no physical or structural resemblance to the referent it denotes.	Isomorphic or homomorphic; the representational medium physically or functionally mirrors the spatial geometry of the referent.
Computational Medium	Abstract, amodal data storage; syntactic engine manipulating symbolic tokens via formal generative rules.	Functionally spatial array (visual buffer); coordinate-space matrix mapping locations to surface properties.
Compositional Syntax	Governed by formal logical grammar, truth-evaluability, and Frege’s compositionality principle.	Governed by spatial geometry, metric distances, continuous topology, and physical boundaries.
Cognitive Penetrability	Highly penetrable; structures can be updated, modified, or reorganized dynamically via beliefs and goals.	Impenetrable operational limits; transformations are structurally constrained by the functional architecture of the buffer.
Transformational Dynamics	Discrete state substitution (e.g., updating an orientation variable from 0 to 180 degrees in constant time).	Continuous analog simulation; must pass through all intermediate trajectory states (e.g., passing sequentially through 45°, 90°, 135°).
Information Granularity	Naturally categorical and abstract; can omit unspecified spatial details (e.g., NEAR(A, B) without exact distance).	Inherently specific; depicting an object requires committing to a specific scale, orientation, location, and visual resolution.
Modality Alignment	Amodal; completely independent of sensory channels; identical code handles vision, audition, touch, and logic.	Modal; inextricably tied to visual and spatial sensorimotor processing mechanisms.
Primary Neural Substrates	Higher-order multimodal associative cortex, prefrontal cortex, left-hemisphere language networks.	Retinotopically organized early visual cortices (V1, V2, V3), parietal spatial pathways, ventral visual stream.
Role of Conscious Phenomenology	Epiphenomenal; the sensation of visual imagination is an idle readout resembling the steam whistle of a train.	Functionally causally efficacious; depictive spatial structures directly enable and compute cognitive solutions.
Neuropsychological Vulnerability	Impaired by semantic dementias, broad intellectual deficits, and left-hemisphere symbolic aphasias.	Impaired by localized occipital lesions, visual agnosias, cortical blindness, and hemispatial visual neglect.

4.1 Structural Format and Representational Medium

The structural divergence between these two systems begins with the basic relationship between the representational token and its external referent. In a propositional framework, the relationship is entirely arbitrary. The string of ASCII characters or neural firing pattern corresponding to DOG has no four-leggedness, no fur, and no biological shape. Just as the word “cat” in English, “chat” in French, and “gato” in Spanish all denote the exact same feline entity through disparate arbitrary phonemes, propositional tokens bear no structural resemblance to their real-world referents. The semantic binding is achieved through a conceptual dictionary or semantic ontology that maps symbolic tokens to meaning networks.

Conversely, in an analog representation, the relationship is structurally isomorphic or homomorphic. An analog representation preserves the topological invariants and metric proportions of the source object. The distance between internal point $A$ and internal point $B$ in the visual buffer is directly proportional to the physical distance between surface feature $A$ and surface feature $B$ on the actual object. If point $A$ is topologically enclosed within contour $C$, the internal coordinate representation similarly encloses point $A$ within contour $C$. There is no arbitrary code that must be decoded; the spatial properties of the medium itself serve as the representation.

This difference dictates the computational syntax of the two frameworks. Propositional structures rely on an unambiguous, formal grammar where predicates govern arguments. A propositional expression such as ABOVE(BALL, BOX) cannot be misread as ABOVE(BOX, BALL) because the syntactic order strictly determines the semantic truth-conditions. In contrast, an analog representation possesses an emergent metric space. If an analog representation places a circle inside a square, and that square inside a triangle, the spatial relationship INSIDE(CIRCLE, TRIANGLE) is automatically and intrinsically represented without needing an explicit symbolic rule to assert it. The spatial medium computes transitivity geometrically rather than logically.

4.2 Cognitive Penetrability and Transformational Mechanics

The concept of cognitive penetrability, formalized by Pylyshyn, serves as the primary diagnostic tool for determining whether a cognitive process is governed by basic functional architecture or by inferential, knowledge-driven operations. A process is cognitively penetrable if an agent’s beliefs, goals, or knowledge can fundamentally alter its operational steps and execution constraints. Propositional systems are thoroughly cognitively penetrable: if a subject learns a new fact or adopts a new problem-solving strategy, their semantic network updates immediately, radically altering the retrieval path and inferential output of the system.

In contrast, the analog transformational engine is hypothesized to possess an impenetrable core governed by intrinsic physical simulation laws. When an individual mentally rotates a mental shape, they cannot simply choose to make the rotation instantaneous while preserving the vivid depictive fidelity of the process. The representational medium is constrained by operational bounds that require intermediate topological states to be instantiated. The computational cost of an analog transformation scales with the physical extent of the transformation: larger angles require longer rotation times, greater distances require longer scanning times, and finer visual details demand continuous magnification operations.

In computational complexity terms, this marks the fundamental difference between step-wise syntactic substitution and intermediate topological trajectory tracing. In a propositional system, transforming an entity requires updating its state variable via a logical substitution operation—a computational process that takes $\mathcal{O}(1)$ time in an indexed memory address or $\mathcal{O}(\log N)$ in a semantic graph. In an analog system, transforming an entity requires traversing a continuous trajectory across a multi-dimensional coordinate space—a process that scales directly with the distance traversed, exhibiting $\mathcal{O}(D)$ or $\mathcal{O}(\theta)$ complexity.

4.3 Modality Specificity and Neural Realization

The debate over representational format inevitably maps onto the neurobiological dichotomy between amodal and modal brain architectures. Propositional theory is fundamentally amodal. It posits that the brain abstracts away sensory-specific inputs, translating visual, auditory, and somatosensory information into a common, universal computational dialect. This dialect is housed primarily in high-order associative cortices, semantic hubs (such as the anterior temporal lobes), and linguistic-logical circuits within the left prefrontal and temporoparietal cortices. In this framework, sensory systems are mere peripheral transducers—input devices that translate physical energy into symbolic Mentalese.

The analog framework is fundamentally modal and sensorimotor-coupled. It asserts that mental representation directly recruits the same cortical sensory machinery that processes primary perceptual inputs. When visualizing, the brain utilizes early retinotopic visual cortices; when imagining an auditory melody, it deploys tonotopic auditory cortices in Heschl’s gyrus; when simulating motor action, it engages the motor, premotor, and supplementary motor areas. The representation is not an abstract translation; it is a top-down, internally generated perceptual reactivation.

This modal-amodal divide is validated by double dissociations in clinical neuropsychology. Patients with profound damage to bilateral temporal semantic networks (such as in semantic dementia) frequently lose the abstract conceptual knowledge of what an animal is, failing to match the word “tiger” to its abstract taxonomic category. Yet, these same patients can sometimes draw a visually accurate picture of a tiger from immediate perceptual recall or navigate complex spatial mazes. Conversely, patients with focal occipitoparietal lesions can retain abstract conceptual knowledge—they know that a tiger is a carnivorous feline possessing striped fur—yet they remain completely unable to visualize the stripes, mentally rotate a simple shape, or perform metric judgments regarding the animal’s physical contours.

5. Similarities and Theoretical Overlap

5.1 Common Functional Objectives in Cognition

Despite their diametric opposition regarding the computational medium of thought, Propositional Theory and the Analog Framework share foundational goals within cognitive psychology. Both theoretical paradigms were formulated to solve the identical scientific problem: explaining how an autonomous cognitive agent internally encodes, stores, manipulates, and retrieves knowledge about the external world in order to generate adaptive, intelligent behavior.

Both models explicitly reject radical behaviorism and operate squarely within the paradigm of representationalism. They concur that human behavior cannot be accounted for by simple stimulus-response contingencies; mediation by internal, structured representations is computationally indispensable. Furthermore, both paradigms share the goal of algorithmic predictability: they formulate precise quantitative hypotheses regarding behavioral response times, error distributions, and cognitive performance profiles. Whether calculating the milliseconds required to traverse an imagined map or the milliseconds required to resolve a categorical syllogism, both frameworks seek to establish a deterministic, mathematically grounded model of human information processing that respects the finite resource constraints of biological brains.

5.2 Dual-Coding and Hybrid Computational Models

As the imagery debate matured, many cognitive scientists recognized that the human mind rarely operates in absolute extremes. Real-world human cognition is neither purely symbolic nor purely pictorial; it is a synergistic, multi-tiered enterprise. This realization gave rise to influential dual-coding and hybrid architectures.

The most celebrated early synthesis was Allan Paivio’s Dual-Coding Theory (DCT), formulated in 1971 and expanded over subsequent decades. Paivio proposed that human memory and cognition comprise two functionally independent yet structurally interconnected representational systems:

• The Nonverbal (Analogue/Depictive) System: Operates via structural structural units termed imagoes. Imagoes retain perceptual and spatial properties, organized as part-to-whole spatial configurations.
• The Verbal (Propositional/Symbolic) System: Operates via discrete structural units termed logogens. Logogens are abstract, language-like symbolic structures organized into sequential, hierarchical linguistic chains.

Paivio demonstrated that memory performance is dramatically enhanced when information is encoded simultaneously through both channels—a phenomenon known as the dual-coding advantage or picture superiority effect. For instance, remembering a concrete noun (such as “apple”) is far more robust than remembering an abstract noun (such as “justice”) because the concrete noun automatically activates both an imago (its visual shape, color, and texture) and a logogen (its linguistic label), creating redundant associative pathways that protect against memory decay and retrieval failure.

In modern computational cognitive science, this synthesis is embodied in contemporary cognitive architectures such as ACT-R (Adaptive Control of Thought—Rational), developed by John R. Anderson, and SOAR, pioneered by John Laird, Allen Newell, and Paul Rosenbloom. These architectures integrate discrete, symbolic production systems (composed of condition-action rules and declarative propositional memory chunks) with continuous spatial and perceptual simulation buffers. When ACT-R solves a complex spatial-reasoning task, its symbolic module executes abstract production rules, while its specialized visuospatial module manipulates metric coordinate frames. Complex, real-world human intelligence relies upon an unbroken, bidirectional feedback loop between abstract logical deduction and analog spatial simulation.

6. When to Apply Propositional vs. Analog Paradigms

6.1 Domain-Specific Cognitive Tasks

Because propositional and analog formats possess distinct computational properties, each format excels at specific classes of cognitive problems while struggling with others. Applying the correct representational paradigm depends on the structural constraints of the cognitive task under examination.

Propositional Paradigms excel in:

• Formal Logical Deduction and Syllogistic Reasoning: Operations involving classical boolean logic, such as negation, disjunction, and material implication, cannot be cleanly represented depictively. One cannot construct an analog depiction of the concept “It is not the case that the circle is red” without depicting the circle as some specific alternative color, which erroneously introduces unasserted metric information. Propositional logic handles negation effortlessly via symbolic operators (e.g., $\neg \text{Red}(\text{Circle})$).
• Abstract Categorization and Taxonomic Inference: Grouping disparate objects under abstract categorical headers—such as classifying a whale, a bat, and a human as MAMMAL—relies on propositional attribute matrices. An analog depiction of a “generic mammal” is mathematically impossible, as George Berkeley pointed out in his classical critique of John Locke’s abstract ideas.
• Counterfactual and Modal Reasoning: Evaluating counterfactual propositions (e.g., “If the train had arrived on time, we would have avoided the traffic”) requires manipulating symbolic possible-world scenarios that have no physical or perceptual reality to simulate directly.

Analog Paradigms excel in:

• Spatial Navigation and Path Integration: Determining the optimal route through an unmapped or geometrically complex environment requires maintaining continuous topological vectors and calculating dynamic self-to-environment spatial coordinates.
• Mechanical Simulation and Kinematic Reasoning: Mary Hegarty’s research into mechanical problem-solving demonstrates that when humans predict the motion of interconnected pulley systems or interlocking gear assemblies, they deploy continuous mental animation. They simulate the physical friction, rotation, and force transfer sequentially through analog spatial models rather than evaluating abstract propositional physics formulas.
• Complex Geometric and Morphological Synthesis: Architects, sculptors, and structural engineers construct mental models where millimeter-level spatial tolerances, balance, and aesthetic composition must be evaluated simultaneously. Slicing an imagined solid cone at an oblique angle to reveal an elliptical cross-section is achieved seamlessly within an analog coordinate buffer, whereas deriving the same conclusion from propositional coordinate geometry requires pages of algebraic calculations.

6.2 Implications for Artificial Intelligence and Robotics

The dichotomy between propositional and analog representation directly parallels the historical and contemporary divisions within Artificial Intelligence (AI) and robotics. The trajectory of AI development can be understood as an ongoing oscillation between these two computational philosophies.

Classical Symbolic Artificial Intelligence—often referred to as GOFAI (Good Old-Fashioned AI)—was the direct engineering manifestation of Propositional Theory. Systems constructed within this paradigm relied upon knowledge graphs, formal ontology languages (such as OWL), expert system rule-bases, and classical first-order predicate logic engines. These architectures stored vast repositories of propositional statements and executed automated theorem proving to derive new truths. While remarkably successful at structured, closed-world problems (such as playing chess or calculating tax compliance), classical symbolic AI suffered catastrophic failure when confronted with the open, continuous, messy physical world—a crisis formalised as the symbol grounding problem (Stevan Harnad) and the frame problem (Daniel Dennett).

Conversely, the fields of Spatial Computing, Computer Vision, and Autonomous Robotics have increasingly embraced analog, depictive principles. Modern robotic navigation rarely relies on a pure database of abstract sentences; it operates through Simultaneous Localization and Mapping (SLAM), continuous coordinate grids, occupancy voxel maps, and distance-field arrays. These systems maintain analog metric representations of the local environment to compute spatial trajectories, collision margins, and visual affordances in real time.

The contemporary paradigm of Neurosymbolic AI represents the state-of-the-art convergence of both worlds. Modern foundation models, such as large language models (LLMs) and multi-modal vision-language transformers, utilize vast continuous mathematical spaces—high-dimensional vector embeddings—to represent meaning. Within these latent embedding spaces, semantic relationships often mirror analog geometries: the vector distance and directional offsets directly encode conceptual relationships (the classic analog vector arithmetic: $\vec{v}_{\text{King}} – \vec{v}_{\text{Man}} + \vec{v}_{\text{Woman}} \approx \vec{v}_{\text{Queen}}$). By fusing continuous geometric tensor representations with higher-order symbolic control modules, neurosymbolic artificial intelligence seeks to resolve the exact tension that ignited the imagery debate fifty years ago.

7. Frequently Asked Questions

7.1 Is mental imagery definitively proven to be non-epiphenomenal?

The balance of contemporary neuroscientific evidence strongly indicates that mental imagery is not an epiphenomenon, but a causally efficacious computational mechanism within the human brain. The definitive proof came through the application of interventionist causal techniques, specifically repetitive Transcranial Magnetic Stimulation (rTMS). When researchers (such as Kosslyn et al., 1999) temporarily disrupted neural processing within retinotopically mapped primary visual area V1 using focused magnetic pulses, human subjects experienced measurable, statistically significant deficits in their ability to perform high-resolution mental imagery tasks. If imagery were purely epiphenomenal—a harmless experiential byproduct like the steam whistle on a train—disrupting the early visual cortex would have no impact on the successful execution of the underlying cognitive computation.

However, strict propositional theorists have maintained nuanced counterarguments. They contend that while sensory cortices are certainly recruited during visualization, this activation may be a downstream sensory reconstruction driven by top-down control networks in the prefrontal and parietal lobes. Under this interpretation, the ultimate decision-making engine remains symbolic, and the sensory activation is a temporary visual scratchpad used only when the symbolic engine demands metric interpolation. Nonetheless, because this scratchpad is required to complete the task successfully, modern cognitive neuroscience almost universally views depictive representations as functionally integrated components of human cognition.

7.2 How does propositional theory explain visual memories of complex faces?

To explain the visual representation of intricate, holistic forms such as human faces, Propositional Theory relies upon hierarchical structural descriptions. Rather than storing a photographic or depictive bitmap of a face, a propositional architecture represents the face as a nested tree-structure of discrete attribute-value pairs linked by spatial-relational predicates.

At the highest structural node, the system instantiates FACE(INDIVIDUAL_X). This node branches into sub-propositions defining constituent structural geometry:

• HAS_PART(FACE, NOSE)
• HAS_PART(FACE, EYES)
• HAS_PART(FACE, JAW)

These parts are then structurally bound by relational predicates: BETWEEN(NOSE, LEFT_EYE, RIGHT_EYE) and ABOVE(NOSE, MOUTH). Metric details are integrated not as continuous surfaces, but as discrete quantitative parameters: INTERPUPIL_DISTANCE(EYES, 63_MM), SHAPE(JAW, ANGULAR), and CURVATURE(CHIN, CONVEX).

The primary critique of this propositional explanation is its extreme computational overhead and vulnerability to perceptual bottlenecks. Human visual recognition of faces is famously holistic, ultra-rapid (occurring within 170 milliseconds, as indexed by the N170 ERP component), and remarkably sensitive to second-order spatial relational changes (such as shifting an eye two millimeters outward). Generating, parsing, and searching millions of nested symbolic propositions within a hundred milliseconds strains biological and algorithmic plausibility, leading most vision scientists to conclude that face recognition relies on specialized, quasi-analog, multi-dimensional feature spaces within the fusiform face area (FFA).

7.3 Can modern deep neural networks resolve the imagery debate?

Modern deep neural networks (DNNs) have radically reframed the terms of the imagery debate by presenting a computational architecture that transcends the classical dichotomy between discrete symbols and literal pictorial buffers. In a deep neural network, information is encoded as continuous vector embeddings within high-dimensional latent spaces. These distributed representations are neither classical discrete propositions (they lack explicit symbolic grammar, alphanumeric tokens, and formal logic) nor literal analog pictures (they are high-dimensional vectors, not two-dimensional coordinate canvases).

However, these latent spaces exhibit striking quasi-analog geometric properties. Within the latent space of a diffusion model or a vision transformer, distance and topological direction correspond directly to semantic and physical similarity. The network can perform smooth, continuous interpolations between concepts, effectively “rotating” or “morphing” representations through continuous mathematical manifolds. Consequently, modern AI suggests that the human brain likely executes a form of representation that blends both frameworks: using continuous, multi-dimensional tensor representations that execute analog-like spatial simulations while simultaneously performing high-level structural compositions that mimic the syntactic power of propositional logic.

8. Conclusion: Synthesizing the Architecture of Thought

8.1 Summary of Core Distinctions

The imagery debate has served as one of the most powerful crucibles in the history of cognitive science, compelling researchers to refine the concepts of representation, architecture, and computational execution. At its core, the division remains profound:

Propositional Theory posits an amodal, symbolic mind. It treats thought as an internal language—a discrete syntactic computation operating over abstract, truth-bearing tokens. Its strength lies in its profound explanatory power for formal logic, categorical abstraction, semantic reasoning, and compositional linguistics. It demystifies the phenomenological mirage of the mind’s eye, protecting cognitive science from the philosophical trap of the homunculus fallacy.

Analog Theory posits a modal, depictive mind. It demonstrates that the physical and functional architecture of the brain supports continuous, coordinate-based representations that directly mirror the spatial, metric, and topological structures of the physical universe. Anchored in rigorous behavioral chronometry and verified through functional neuroimaging and cortical lesion studies, the analog framework proves that the brain uses its sensory and motor machinery not merely to perceive the world, but to simulate it internally.

Neither theory alone captures the full spectrum of human mental life. To reduce the human mind exclusively to propositional code is to ignore the visceral, spatial reality of human imagination, the geometric computation of our sensory cortices, and the smooth kinematic simulations that make physical interaction possible. To reduce the mind exclusively to analog pictures is to leave human cognition incapable of negation, abstract categorization, formal deduction, and the unbounded combinatorial power of language.

8.2 The Modern Consensus and Future Horizons

The contemporary consensus within cognitive science and neuroscience has largely moved past the adversarial either/or framing of the twentieth century. The emerging paradigm embraces embodied cognition, perceptual symbol systems (Lawrence Barsalou), and predictive processing (Andy Clark, Karl Friston). Under these modern frameworks, mental representations are understood as grounded sensory simulations nested within hierarchical predictive models.

In this view, the brain is an active, generative prediction machine. When performing abstract reasoning, the mind regularly deploys perceptual simulations—reactivating early modal cortices in an analog fashion to predict the sensory consequences of potential actions or to test geometric hypotheses. Concurrently, higher-order prefrontal and temporal networks enforce abstract structural constraints, orchestrating these modal simulations and translating their continuous outputs into categorical decisions and symbolic actions.

The resolution of the imagery debate is thus fundamentally synthetic. Propositional Theory accurately describes the high-level, symbolic, compositional control architecture of thought, while Analog Representation accurately describes the lower-level, continuous, topographic simulators that ground our thoughts in the physical geometry of the reality we inhabit. Far from being mutually exclusive, the descriptive code and the depictive canvas are complementary computational engines of the human mind, working in continuous harmony to render the marvel of human intelligence possible.

References

• Anderson, J. R. (1983). The Architecture of Cognition. Harvard University Press.
• Barsalou, L. W. (1999). Perceptual symbol systems. Behavioral and Brain Sciences, 22(4), 577-660.
• Block, N. (1981). Imagery. MIT Press.
• Cooper, L. A., & Shepard, R. N. (1973). Chronometric studies of the rotation of mental images. In W. G. Chase (Ed.), Visual Information Processing (pp. 75-176). Academic Press.
• Farah, M. J. (1984). The neurological basis of mental imagery: A componential analysis. Cognition, 18(1-3), 245-272.
• Fodor, J. A. (1975). The Language of Thought. Harvard University Press.
• Hegarty, M. (2004). Mechanical reasoning by mental simulation. Trends in Cognitive Sciences, 8(6), 280-285.
• Kosslyn, S. M. (1973). Scanning visual images: Some structural implications. Perception & Psychophysics, 14(1), 90-94.
• Kosslyn, S. M. (1975). Information representation in visual images. Cognitive Psychology, 7(3), 341-370.
• Kosslyn, S. M. (1980). Image and Mind. Harvard University Press.
• Kosslyn, S. M. (1994). Image and Brain: The Resolution of the Imagery Debate. MIT Press.
• Kosslyn, S. M., Ball, T. M., & Reiser, B. J. (1978). Visual images preserve metric spatial information: Evidence from studies of image scanning. Journal of Experimental Psychology: Human Perception and Performance, 4(1), 47-60.
• Kosslyn, S. M., Pascual-Leone, A., Felician, O., Camposano, S., Keenan, J. P., Thompson, W. L., Ganis, G., Sukel, K. E., & Alpert, N. M. (1999). The role of Area 17 in visual imagery: Convergent evidence from PET and rTMS. Science, 284(5411), 167-170.
• Kosslyn, S. M., Thompson, W. L., Kim, I. J., & Alpert, N. M. (1995). Topographical representations of mental images in primary visual cortex. Nature, 378(6556), 496-498.
• Paivio, A. (1971). Imagery and Verbal Processes. Holt, Rinehart and Winston.
• Paivio, A. (1986). Mental Representations: A Dual Coding Approach. Oxford University Press.
• Pylyshyn, Z. W. (1973). What the mind’s eye tells the mind’s brain: A critique of mental imagery. Psychological Bulletin, 80(1), 1-24.
• Pylyshyn, Z. W. (1981). The imagery debate: Analogue media versus tacit knowledge. Psychological Review, 88(1), 16-45.
• Pylyshyn, Z. W. (2002). Mental imagery: In search of a theory. Behavioral and Brain Sciences, 25(2), 157-182.
• Pylyshyn, Z. W. (2003). Seeing and Visualizing: It’s Not What You Think. MIT Press.
• Shepard, R. N., & Metzler, J. (1971). Mental rotation of three-dimensional objects. Science, 171(3972), 701-703.
• Tootell, R. B., Silverman, M. S., Switkes, E., & De Valois, R. L. (1982). Deoxyglucose analysis of retinotopic organization in primate striate cortex. Science, 218(4575), 902-904.