Modular Mind vs. Connectionist Mind

The quest to decipher the architecture of human cognition represents one of the most profound and fiercely contested intellectual undertakings in modern science. At its core, cognitive science seeks to identify the fundamental mechanisms through which the brain receives, transforms, stores, retrieves, and utilizes information to generate thought, language, perception, and deliberate action. Over the past half-century, this inquiry has crystallized into a monumental theoretical schism between two radically different paradigms: the Modular Mind and the Connectionist Mind. Each framework offers a distinct vision not only of how the mind processes information, but also of how mental properties relate to physical brain tissue, how human cognitive capacities evolve and develop, and how artificial computational systems should be engineered to mirror or achieve human-level intelligence.

The Modular Mind framework treats cognition as an orchestrated collection of discrete, specialized, and semi-independent computational sub-mechanisms or “modules.” Rooted historically in the classical computational theory of mind, rationalist epistemology, and evolutionary biology, modularity posits that the human mind is akin to a complex, multi-component engineering triumph. In this paradigm, distinct cognitive tasks—from parsing the syntax of an utterance to recognizing a conspecific face or identifying spatial boundaries—are governed by dedicated, encapsulated information-processing units. These units operate autonomously, insulated from the vast, holistic web of general beliefs and desires that populate central awareness. By compartmentalizing computational tasks, the modular approach explains the speed, mandatory execution, and functional resilience of human perceptual and cognitive performance, as well as the striking selective deficits observed following focal brain injury.

In direct opposition stands the Connectionist Mind framework, known colloquially as Parallel Distributed Processing (PDP) or neural network modeling. Inspired by the dense, parallel, and continuous neurobiology of the cerebral cortex, connectionism rejects the notion of rigid, pre-packaged, domain-specific computational boxes. Instead, connectionists model the mind as an expansive, interconnected web of simple, neuron-like processing units whose collective, continuous interactions give rise to emergent cognitive behaviors. Within connectionism, knowledge is neither stored in isolated symbolic addresses nor governed by explicit, prescriptive rules. Rather, representations are distributed as high-dimensional activation vectors across extensive neural populations, and cognitive operations unfold through continuous, statistical adjustments of synaptic weights. By viewing cognition through the lens of dynamic self-organization, continuous learning, and biological plausibility, connectionism shifts the explanatory focus from pre-configured formal structures to adaptive plasticity, statistical induction, and graceful degradation.

1. Introduction: The Great Architectural Debate in Cognitive Science

1.1 Defining the Core Frameworks

To understand the profound tension between the modular and connectionist views, one must first delineate the operational principles that define each paradigm. The Modular Mind models cognition as an ensemble of specialized, autonomous, and structurally encapsulated computational modules. In this framework, human cognitive architecture is not an undifferentiated, general-purpose processing engine; rather, it is functionally carved at its joints into discrete sub-components. Each module is optimized to handle a restricted class of inputs, executing algorithmic transformations governed by internal, domain-specific constraints. The core claim is that cognition can be cleanly decomposed into distinct functional components whose internal computational dynamics run independently of the wider cognitive ecosystem. The operational signature of a module is functional isolation: its inner workings proceed without drawing upon general world knowledge, central goals, or the simultaneous operations of neighboring modules.

Conversely, the Connectionist Mind conceives cognition as emergent patterns that self-organize across vast, interconnected, subsymbolic neural networks. Rather than routing distinct categories of information through isolated computational conduits, connectionism posits a unified processing substrate composed of nodes (idealized processing units) and directed, weighted connections (idealized synapses). Information processing in a connectionist system is intrinsically parallel and widely distributed: an input does not activate a dedicated, circumscribed sub-program, but instead perturbs the overall state of the network, triggering a cascade of continuous activation updates across layers of units until the system settles into an energetic minimum or stable attractor state representing a cognitive decision, interpretation, or motor plan.

Both theories share the overarching goal of explaining mental operations, representational structures, and the physical realization of thought within physical substrates. Yet their descriptive primitives diverge completely. Where modularity identifies discrete symbols, explicit syntactical rules, and bounded algorithmic pathways, connectionism identifies vector spaces, continuous numerical activations, non-linear transformation functions, and multidimensional weight matrices. Consequently, the debate is not merely an empirical disagreement over brain mapping; it is a fundamental epistemological dispute concerning the nature of mental representation and the correct computational level of analysis required to explain the mind.

1.2 Historical Emergence of the Debate

The philosophical and scientific roots of this architectural divide can be traced back to historical tensions between rationalist and empiricist conceptions of the mind, which were subsequently revitalized by twentieth-century computational revolutions. The modular paradigm found its formal twentieth-century expression in the classical symbolic cognitive science pioneered by Alan Turing, Noam Chomsky, Herbert Simon, and Allen Newell. This tradition viewed the mind as a physical symbol system that operates via the formal, syntactic manipulation of discrete mental tokens. The definitive philosophical manifesto of this perspective emerged with Jerry Fodor’s groundbreaking 1983 work, The Modularity of Mind. Fodor synthesized linguistics, phrenological history, and cognitive psychology to articulate an architecture featuring encapsulated, fast, mandatory input analyzers that feed a non-modular, holistic central processor.

Simultaneously, an intellectual counter-revolution was gathering momentum. Dismayed by the biological implausibility, architectural fragility, and computational bottlenecks of classical symbolic systems—often referred to pejoratively as “Good Old-Fashioned Artificial Intelligence” (GOFAI)—a coalition of psychologists, physicists, and computer scientists sought inspiration directly from neuroanatomy. Although neural modeling had early precursors in the work of Warren McCulloch, Walter Pitts, and Frank Rosenblatt’s Perceptron, the framework exploded into mainstream cognitive science with the publication of the seminal two-volume work, Parallel Distributed Processing: Explorations in the Microstructure of Cognition, authored by David Rumelhart, James McClelland, and the PDP Research Group in 1986. Rumelhart and McClelland demonstrated that multi-layered networks utilizing distributed representations and learning algorithms like backpropagation could learn complex cognitive mappings—such as the formation of the English past tense—without needing pre-programmed, symbolic rules.

The publication of the PDP manifesto catalyzed a fiery intellectual schism. The debate was formally initiated when Jerry Fodor and Zenon Pylyshyn published their 1988 critique, “Connectionism and cognitive architecture: A critical analysis.” Fodor and Pylyshyn argued that connectionist networks, lacking an explicit combinatorial syntax and compositional semantics, were fundamentally incapable of explaining the core properties of human thought: systematicity, productivity, and inferential coherence. In response, connectionists countered that classical modular symbol systems were brittle abstractions that failed to account for real-time biological learning, noise tolerance, developmental plasticity, and the continuous, context-sensitive nature of human experience.

1.3 Significance in Modern Cognitive Science and Artificial Intelligence

Far from being a resolved historical curiosity, the modular-versus-connectionist debate remains the central architectural axis around which contemporary cognitive science, neuropsychology, evolutionary psychology, and artificial intelligence rotate. In cognitive psychology and evolutionary theory, the debate manifests in disagreements over how human intelligence evolved. Evolutionary psychologists, taking modularity to its theoretical extreme, have advanced the “Massive Modularity Hypothesis,” asserting that natural selection must have engineered a mind composed entirely of hundreds or thousands of specialized, adaptive modules—a cognitive “Swiss Army knife”—designed to resolve ancestral survival challenges ranging from predator avoidance to social exchange.

In modern neuroscience and neuropsychology, the debate directly informs how empirical data are gathered, analyzed, and interpreted. When cognitive neuroscientists use functional magnetic resonance imaging (fMRI) or magnetoencephalography (MEG) to observe brain activity, their interpretations are deeply shaped by these theoretical commitments. A researcher working within a modular paradigm looks for discrete, specialized brain regions dedicated exclusively to specific functions, interpreting localized hemodynamic spikes as evidence for distinct neural modules. Conversely, a connectionist or network neuroscientist interprets the same data through the lens of large-scale distributed functional networks, treating localized activations merely as integrated convergence zones within a dynamically shifting, brain-wide computational web.

Crucially, the debate has become the defining ideological battleground of contemporary Artificial Intelligence (AI) and the quest for Artificial General Intelligence (AGI). The spectacular real-world success of modern deep learning—which represents the direct computational lineage of 1980s connectionism—demonstrates that massive, end-to-end connectionist architectures trained on vast datasets can master complex visual perception, natural language translation, and strategic reasoning without requiring hand-engineered, domain-specific modules. However, these pure connectionist systems struggle with hallucination, a lack of causal reasoning, out-of-distribution generalization, and explainability. Consequently, leading theorists now debate whether true artificial general intelligence can be achieved simply by scaling up continuous connectionist neural networks, or whether it will ultimately require integrating structured, symbolic, and functionally encapsulated modular components to enforce logical consistency and rule-governed inference.

2. The Modular Mind: Architectural Foundations and Core Tenets

2.1 Fodorian Modularity and Structural Encapsulation

The classical architectural foundation of modularity was articulated with rigorous philosophical precision by Jerry Fodor in his 1983 monograph. Fodor did not claim that the entire mind was modular; rather, he drew a fundamental structural line between what he termed input systems (perceptual mechanisms and low-level linguistic parsers) and central systems (mechanisms responsible for reasoning, belief fixation, problem-solving, and general decision-making). According to Fodor, input systems are modular computational devices designed to rapidly deliver veridical representations of the immediate environment to the central cognitive mechanisms, which then deliberate, integrate cross-modal information, and decide upon action.

Fodor established a suite of diagnostic criteria that characterize a true cognitive module. Among these properties, the absolute theoretical cornerstone is informational encapsulation. Informational encapsulation dictates that the internal operations of a module are entirely isolated from the beliefs, desires, goals, and higher-order representations housed within the rest of the mind. The processing within an encapsulated module is purely bottom-up and data-driven; it cannot be penetrated, guided, or overridden by central epistemic states. The quintessential proof of informational encapsulation is the persistence of cognitive and perceptual illusions, such as the famous Müller-Lyer illusion. Even when an individual takes a ruler, measures both lines, and forms a firm, mathematically justified, central belief that the two lines are precisely identical in length, their visual input module continues to perceive one line as longer than the other. The cognitive module responsible for spatial-visual parsing simply cannot access or incorporate the central belief system’s explicit knowledge.

In addition to informational encapsulation, Fodorian modules possess several other structural properties:

• Domain Specificity: A module operates exclusively on a highly restricted, dedicated class of environmental inputs. A phonological decoding module responds solely to acoustic speech waveforms and is blind to visual input, olfactory signals, or non-linguistic auditory frequencies.
• Mandatory Execution: The operations of a module are automatic and cognitively involuntary. An agent cannot choose not to hear an utterance in their native tongue as linguistic speech, nor can they choose to view a recognized face as an unrecognizable array of pixels. The module processes the input as an automatic reflex.
• Fast Processing: Because modules are encapsulated and domain-specific, they do not waste computational time querying the broader web of an agent’s global knowledge base; they compute their specific algorithms nearly instantaneously.
• Shallow Outputs: The representations output by a module are simple, basic-level perceptual categorizations (e.g., identifying an object as a “dog” rather than an “expensive golden retriever belonging to my neighbor”), leaving nuanced semantic and contextual integration to central systems.
• Fixed Neural Architecture and Specific Breakdown Patterns: Modules are presumed to map onto dedicated, localized neurological pathways, meaning that localized trauma results in selective, characteristic cognitive deficits while sparing surrounding operations.

Fodor was famously pessimistic regarding the scientific tractability of central cognitive systems. Because central systems are responsible for holistic belief fixation—which Fodor characterized as isotropic (any belief can be relevant to any other belief) and Quinean (the confirmation of a belief depends on the epistemic status of the entire belief system)—they lack informational encapsulation. Consequently, Fodor formulated his “First Law of the Nonexistence of Cognitive Science,” which maintained that the more holistic and unencapsulated a mental process is, the less anyone understands it, and that true scientific computational modeling could only succeed when directed at encapsulated, peripheral input modules.

2.2 Massive Modularity and Evolutionary Adaptation

While Fodor restricted modularity strictly to peripheral sensory processors and low-level linguistic analysis, a subsequent wave of cognitive scientists, evolutionary biologists, and philosophers expanded the concept into an all-encompassing paradigm known as the Massive Modularity Hypothesis (MMH). Championed by scholars such as Leda Cosmides, John Tooby, Dan Sperber, and Peter Carruthers, massive modularity rejects Fodor’s partition between modular input systems and a non-modular central processor. Instead, MMH posits that the mind is modular “all the way down” and “all the way up”—a dense constellation of hundreds or thousands of evolved, genetically canalized computational programs designed by natural selection to solve specific adaptive problems faced by our hominin ancestors during the Pleistocene epoch.

To substantiate this claim, evolutionary psychologists invoked the metaphor of the cognitive “Swiss Army knife.” A Swiss Army knife does not accomplish its utility via a single, general-purpose blade; rather, it contains a diverse array of specialized tools—a corkscrew, a scissors, a magnifying glass, a knife blade—each structurally configured to resolve an independent task. Cosmides and Tooby argued that natural selection could not have evolved a purely domain-general cognitive architecture because domain-general systems are computationally intractable when confronted with the complex demands of ancestral environments. They formulated two primary theoretical arguments: the argument from error and the argument from computational tractability (or the avoidance of combinatorial explosion).

The argument from error points out that what counts as adaptive behavior varies completely across distinct biological contexts. For instance, what constitutes a successful behavioral strategy for selecting a mate is fundamentally incompatible with the criteria used for choosing nutritious food, avoiding apex predators, or negotiating social dominance hierarchies. A domain-general mechanism lacking innate, task-specific heuristics would have to evaluate potential actions without any pre-existing evaluative criteria, inevitably leading to fatal errors before learning could occur. A dedicated “food selection module” relies on innate heuristics (e.g., disgust toward bitter or rotting organic substances), whereas a “cheater-detection module” relies on social-contract logic.

The computational tractability argument, closely tied to the artificial intelligence frame problem, asserts that a general-purpose, non-modular system tasked with solving an open-ended environmental challenge would suffer from catastrophic combinatorial explosion. If a system must consider every piece of stored knowledge, every possible permutation of causes, and every potential behavioral pathway without encapsulated computational constraints, it will grind to an algorithmic halt. By restricting the scope of information that any given module can access, massive modularity ensures that the computational search space remains small, tractable, and biologically viable. Under this view, human higher reasoning is not an unencapsulated, holistic void; it is a collaborative assembly of specialized modules for theory of mind, kin recognition, coalitional tracking, spatial navigation, and hazard management.

2.3 Neuropsychological Evidence and Double Dissociations

The primary empirical pillar supporting the modular architecture of mind has historically come from clinical neuropsychology, specifically through the observation of focal brain damage and the discovery of double dissociations. A single dissociation occurs when structural trauma to a specific, circumscribed brain area impairs cognitive function A while leaving cognitive function B intact. However, a single dissociation does not definitively prove that functions A and B rely on independent modules; function A might simply be computationally harder than function B, requiring a higher threshold of general neural resources that were partially depleted by the trauma. To decisively demonstrate functional modularity, researchers require a double dissociation: Patient 1 shows damage to area X, exhibiting impaired function A but normal function B; Patient 2 shows damage to area Y, exhibiting normal function A but impaired function B.

A classic illustration of this modular dissociation occurs in the neuropsychology of vision and facial recognition. Patients suffering from acquired prosopagnosia—frequently resulting from localized lesions to the fusiform face area (FFA) within the ventral occipitotemporal cortex—lose the capacity to identify familiar faces. In extreme cases, these individuals cannot recognize their own reflection in a mirror, their spouse, or their children. Yet, their general visual object agnosia is absent: they can instantly identify coffee mugs, cars, houses, and written text. Crucially, the reverse condition has also been documented: patients with visual object agnosia who can recognize human faces with normal speed and accuracy, but cannot identify inanimate objects. The existence of this double dissociation provides powerful empirical support for the claim that facial recognition is not an emergent expression of general visual problem-solving, but rather an anatomically localized, functionally segregated cognitive module.

Similarly, classical aphasiology has long pointed to double dissociations in human language capabilities to argue for modular organization. Damage to Broca’s area (the left posterior inferior frontal gyrus) classically yields non-fluent, agrammatic speech production, where patients retain auditory comprehension but cannot execute the syntactical and motor-articulatory rules required to produce fluent sentences. In stark contrast, damage to Wernicke’s area (the left superior temporal gyrus) produces fluent, phonologically effortless speech that is completely devoid of semantic coherence—a “word salad”—accompanied by profound comprehension deficits. This historical dissociation between the syntactic-articulatory module and the semantic-comprehension module formed the foundational bedrock of modular neuropsychology.

Modern functional neuroimaging technologies, including high-resolution fMRI and positron emission tomography (PET), have further populated the modular catalog. Researchers have identified specialized cortical patches that respond with high selectivity to specific environmental stimuli: the Parahippocampal Place Area (PPA) for scenes and landscapes, the Extrastriate Body Area (EBA) for human bodies, and the Visual Word Form Area (VWFA) for orthographic patterns. For modular theorists, these findings converge to form an unassailable empirical truth: the brain does not operate as an undifferentiated, holistic organ, but rather partitions computational labor into discrete, anatomically identifiable, encapsulated functional processors.

3. The Connectionist Mind: Parallel Distributed Processing and Emergence

3.1 Subsymbolic Processing and Distributed Representations

While the modular paradigm draws its theoretical inspiration from digital computers and formal logic, the Connectionist Mind grounds its conceptual architecture in the structural principles of neurobiology. Rather than parsing the mind into high-level, domain-specific functional boxes, connectionism treats cognition as the natural, emergent consequence of parallel distributed processing occurring across large networks of simple computational elements known as nodes or artificial neurons. In this framework, the fundamental level of scientific explanation is shifted downward, from the symbolic to the subsymbolic.

In a classical modular-symbolic model, knowledge is stored as discrete, localized symbols that resemble words in a formal language (e.g., [DOG], [CHAIR], [IS_A_MAMMAL]). These symbols have explicit spatial addresses in memory, possess an unalterable syntactic identity, and are manipulated by external, centralized central processing units (CPUs) that apply explicit rules. Connectionism categorically abolishes this separation between processor and data storage, as well as the concept of the localized, atomic symbol. Within a connectionist network, an individual node does not correspond to a high-level cognitive concept. A single node does not mean “dog” or “grandmother.” Rather, a concept is represented as a specific, high-dimensional distributed pattern of activation across an entire population of interconnected nodes.

To conceptualize this, consider a connectionist vector space. If a hidden layer consists of 1,000 artificial neurons, any given mental representation is simply a vector of 1,000 real-valued numbers, each representing the instantaneous activation level of a single unit (typically normalized between 0.0 and 1.0, or -1.0 and +1.0). The concept of a “dog” is realized when the network enters a specific vector state $\vec{v}_1 = \langle 0.12, -0.85, 0.44, dots \rangle$; the concept of a “cat” is represented by another vector $\vec{v}_2 = \langle 0.15, -0.79, 0.41, dots \rangle$. Because semantic representations exist as points in a continuous, multidimensional geometric space, semantic similarity is naturally formalized as geometric proximity (measured via Euclidean distance or cosine similarity). Similar concepts sit close to one another in activation space, allowing the network to exhibit spontaneous inductive generalization: knowledge gained about “dogs” automatically generalizes to “cats” because their corresponding vectors occupy adjacent computational neighborhoods, without requiring a centralized rule-based deduction.

Crucially, this architecture eliminates the conventional computer science dichotomy between “hardware” and “software.” A connectionist network does not possess a fixed physical machine executing an arbitrary, downloadable software program. The “knowledge” of the network resides entirely within the configuration of its connection weights—the numerical values that scale the strength of signals transmitted between nodes, conceptually analogous to biological synaptic strengths. The software is the hardware; the memory is the processor. Processing occurs as incoming numerical activations pass through weight matrices, are summed by target units, and are converted via non-linear mathematical activation functions (such as the sigmoid, hyperbolic tangent, or rectified linear unit [ReLU]) to yield new patterns of activation. Cognition is not the sequential execution of symbolic instructions; it is the continuous relaxation of a dynamic physical system through a complex landscape of activation energy.

3.2 Learning Mechanisms and Synaptic Plasticity

One of the most consequential departures connectionism makes from the modular paradigm concerns ontogeny and learning. The modular approach, particularly in its massive modularity formulation, leans heavily toward genetic nativism: because complex, encapsulated modules could hardly coalesce spontaneously from undifferentiated inputs, they are presumed to be largely hardwired into the genome via millions of years of evolutionary adaptation. Connectionism, by contrast, prioritizes adaptive epigenetic plasticity and statistical learning. It views the structure of the mind not as a rigid blueprint dictated by DNA, but as an emergent developmental trajectory carved by the continuous statistical regularities embedded within environmental interactions.

Connectionist networks learn by modifying their synaptic connection weights over time through mathematically formulated learning rules. The earliest and most biologically transparent of these is Hebbian learning, derived from Donald Hebb’s 1949 neurophysiological postulate that “neurons that fire together, wire together.” In a purely Hebbian framework, if node $A$ and node $B$ are simultaneously active, the connection weight between them increases proportionally:
$$\Delta w_{ij} = \eta \cdot a_i \cdot a_j$$
where $eta$ represents a learning rate parameter, and $a_i$ and $a_j$ represent the activation levels of the presynaptic and postsynaptic units, respectively. Hebbian learning operates in an unsupervised manner, allowing connectionist networks to discover associative structures, statistical correlations, and clustering patterns inherent in environmental inputs without external supervision.

For complex cognitive tasks involving explicit mappings between ambiguous sensory inputs and complex outputs (such as language acquisition), connectionists developed error-correcting, supervised learning algorithms, most notably the backpropagation of error (Rumelhart, Hinton, & Williams, 1986). In backpropagation, a network is presented with an input pattern, which propagates forward through successive hidden layers to produce an output vector. This observed output is compared against a target vector to compute an objective error function (often mean squared error or cross-entropy). The algorithm then calculates the partial derivative of the error with respect to each individual weight in the network via the calculus chain rule:
$$\frac{\partial E}{\partial w_{ij}} = \frac{\partial E}{\partial net_j} \cdot \frac{\partial net_j}{\partial w_{ij}}$$
These gradients are propagated backward through the network, incrementally shifting every weight in the opposite direction of the gradient via gradient descent. Through millions of micro-adjustments, the network continuously sculpts its internal multidimensional energy landscape, autonomously discovering intermediate representations in its hidden layers that capture the deep, latent statistical regularities of the training environment. Rules are never explicitly written; rather, rule-like behavior emerges naturally from the optimization of continuous statistical weights.

3.3 Biological Realism, Fault Tolerance, and Graceful Degradation

A primary justification for the connectionist framework has always been its deliberate fidelity to the structural and computational realities of biological nervous systems. The classical, modular, symbolic model relies on an architecture that looks remarkably like a von Neumann digital computer: a central processing unit, sequential serial processing cycles, discrete memory addresses, and brittle symbolic pointers. However, real mammalian brains share virtually none of these structural traits. The cerebral cortex contains roughly 86 billion neurons, each connected to thousands of others via an estimated 100 trillion synapses. Communication is fundamentally asynchronous, parallel, continuous, and noisy. Connectionism embraces these neurobiological constraints as essential computational virtues rather than inconvenient biological bugs.

A direct, profound consequence of distributed connectionist architecture is graceful degradation. In a classical modular or symbolic computer system, information storage is highly localized and brittle. If a single memory address containing a pointer is corrupted, or if an essential line of code in an encapsulated module’s algorithm is deleted, the entire system often crashes catastrophically. The program produces a fatal system error, halting all processing. This brittleness stands in glaring contradiction to the resilience of biological brains, which continuously withstand the daily loss of thousands of neurons due to natural attrition, vascular micro-accidents, or physical trauma without suddenly losing the capacity to speak, perceive, or reason.

Connectionist networks exhibit graceful degradation precisely because their knowledge is distributed diffusely across an entire network of weights. If an experimenter simulates physical brain injury by randomly pruning 10%, 20%, or even 40% of the weights in a trained neural network, or by injecting continuous Gaussian noise into unit activations, the network does not abruptly fail. Instead, its performance degrades slowly, continuously, and gracefully. The network’s output vectors become slightly less precise, its reaction times slow, and its classifications may demonstrate mild ambiguity, but its core functional competence remains intact. The system retains a high tolerance for noisy, incomplete, or partially corrupted inputs.

Furthermore, this architectural tolerance naturally yields spontaneous pattern completion and content-addressable memory. In a modular, address-based memory system, an agent can only retrieve an item by knowing its exact physical memory address or evaluating a precise logical search query. In a connectionist network, presenting a noisy, incomplete fragment of a familiar input pattern automatically drives the dynamic state of the network down its familiar energy gradient toward the nearest established attractor basin. The network spontaneously reconstructs the missing features, filling in the blanks based on the statistical probabilities embedded in its weights—a phenomenon that mirrors human associative memory with remarkable accuracy.

4. Key Differences Between Modular and Connectionist Architectures

The differences between modular and connectionist paradigms span every significant dimension of cognitive science, from high-level epistemology down to low-level computational mechanics. The following head-to-head comparison table systematically contrasts these two foundational architectures across their most critical theoretical and operational dimensions:

Dimension / Aspect	The Modular Mind	The Connectionist Mind
Fundamental Computational Unit	Discrete, domain-specific modules executing formal algorithms on localized symbols.	Interconnected artificial neurons (nodes) adjusting scalar activation levels.
Representational Format	Explicit, symbolic, language-like structures with formal, compositional syntax (Language of Thought).	Distributed, subsymbolic, real-valued high-dimensional vectors across continuous activation spaces.
Information Processing Flow	Rigid, compartmentalized pipelines; serial or strictly staged processing with encapsulated boundaries.	Massively parallel, continuous, distributed signal propagation across recurrent or feedforward layers.
Epistemic Encapsulation	High: internal operations of a module are insulated from external central beliefs and cross-modal data.	Low to Non-existent: continuous global interconnectivity; activations can propagate across the entire network.
Ontogeny and Origin	Innate, canalized, and genetically pre-programmed through evolutionary natural selection.	Plastic, experiential, and developmental; representations emerge via continuous statistical learning.
Mechanism of Learning	Parameter setting, algorithmic triggering, or innate maturation; limited role for empiricist weight shifting.	Hebbian learning, error-driven delta rules, backpropagation, and continuous gradient descent.
Hardware-Software Relation	Strict dualism: cognitive software (rules and symbols) is computationally independent of physical neural wetware.	Monistic embodiment: the program is the physical structure; software is entirely conflated with synaptic connectivity.
Systematicity and Compositionality	Inherent and structural: symbols combine recursively through formal syntactic rules.	Emergent and approximate: vector spaces approximate compositionality through tensor products or learned mappings.
Response to Physical Damage	Clean, localized, domain-specific dissociations; complete loss of single faculties while others remain intact.	Graceful degradation; progressive, diffuse performance loss with high tolerance for noisy or partial inputs.
Handling of Context	Context is strictly bounded and excluded from encapsulated subroutines to avoid computational explosion.	Context is inherently integrated as an omnipresent influence across the continuous multidimensional state space.

4.1 Architectural Organization and Information Flow

The structural topology of the Modular Mind is characterized by discrete, compartmentalized boundaries. Information flow within a modular architecture proceeds through pre-determined, rigid conduits. An environmental stimulus is captured by specialized sensory transducers, converted into an internal format, and immediately routed to a designated, domain-specific module. The boundaries of this module act as an epistemic firewall: external data, central expectations, and alternative interpretations cannot cross the threshold to contaminate the module’s internal computations. Processing is largely feedforward, pipeline-oriented, and staged. Once the module completes its specific algorithmic manipulation, it yields a shallow output that is deposited into the central processing pool. Functional isolation is the governing architectural design principle; the system maintains operational integrity through strict segregation of computational labor.

Conversely, the Connectionist Mind is organized as a diffuse, unified, and continuous dynamical system. The architectural topology is defined by dense layers of nodes linked through reciprocal, often recurrent connectivity. Information does not flow down an isolated, private hallway; instead, it ripples across an open, continuous network. The boundaries between “perceptual processing,” “contextual evaluation,” and “decision-making” are thoroughly blurred because activation waves propagate continuously backward and forward throughout the system. Rather than executing compartmentalized algorithms in functional isolation, connectionist architectures rely on global or regional dynamic balance: processing is an act of simultaneous constraint satisfaction, wherein multiple, often conflicting sources of statistical evidence are weighed concurrently until the entire network settles into a harmonious state of minimum energy.

4.2 Representational Formats: Classical Symbols vs. Distributed Activations

At the center of the dispute between these two models lies the representational format used to depict reality. The modular mind relies on what Jerry Fodor famously christened the Language of Thought or Mentalese. In this view, mental representations are physical symbol tokens that possess both syntactic structure and semantic content. Just as natural language sentences derive their meaning from the compositional arrangement of discrete words governed by grammatical rules, thoughts derive their meaning from the systematic, rule-governed arrangement of discrete symbols. The mental representation [JOHN LOVES MARY] is structurally distinct from [MARY LOVES JOHN], yet both are composed of identical, recombined symbolic primitives. This classical formulation easily accounts for the infinite productivity of the human mind (our capacity to generate an unbounded set of novel thoughts from finite primitives) and its systematicity (the intrinsic truth that the ability to think “John loves Mary” necessarily entails the ability to think “Mary loves John”).

Connectionism completely re-envisions this representational paradigm by discarding discrete symbols in favor of high-dimensional, continuous vector representations. In a connectionist network, there are no atomic, indivisible tokens manipulated by an explicit syntax. Instead, concepts are dynamic vectors inhabiting continuous metric spaces. Meaning does not reside within an isolated, discrete component; it is distributed across thousands of non-linearly interacting variables. The challenge for connectionism, as Fodor and Pylyshyn forcefully pointed out in their 1988 critique, has always been accounting for true compositionality and systematicity without smuggling classical symbols in through the back door. Connectionists argue that systematicity in human cognition is rarely absolute or purely formal; rather, it is soft, context-dependent, and analogical. Connectionist systems achieve functional compositionality through sophisticated mathematical techniques—such as circular convolutions, tensor products, or recursive auto-associative memories—which represent complex structural relationships without freezing them into rigid, unyielding syntactic formalisms.

4.3 Ontogeny: Innate Specialization vs. Plastic Development

The two paradigms stand on opposing ends of the classical nature-versus-nurture spectrum regarding the ontogenetic origins of mental faculties. The modular framework is historically aligned with rationalist, nativist traditions. Evolutionary psychologists and strong modularists assert that the structural architecture of modules is fundamentally innate, specified by the human genetic code as the product of natural selection. Under this view, cognitive modules do not need to be painfully derived from scratch through undifferentiated environmental learning; they mature along tightly canalized developmental trajectories, requiring only minimal, triggering environmental inputs to come online. The human infant is born equipped with an innate “starter kit” of dedicated computational machinery: an innate theory of mind module, an innate naive physics module, and an innate language acquisition device (Universal Grammar). Learning, in a modular framework, is largely a process of parameter setting or filling in the localized databases of pre-existing computational engines.

Connectionism, conversely, embodies a modern, computational form of developmental empiricism, championed extensively by developmental cognitive scientists like Jeffrey Elman, Elizabeth Bates, and Mark Johnson in their seminal work, Rethinking Innateness: A Connectionist Perspective on Development. Connectionists argue that what is genetically inherited is not a collection of fully formed, domain-specific modules, but rather a set of global architectural constraints, general computational principles, and sensory-motor biases. The apparent functional specialization observed in adult brains is not an innate starting point; it is the progressive, emergent consequence of self-organizing developmental processes operating over years of dense environmental exposure. Under this view, the brain becomes modularized over developmental time through competitive plasticity, rather than starting as a collection of pre-packaged, hardwired modules. The mind’s architecture is shaped through constant, dynamic bidirectional transactions between plastic neural networks and rich, structured sensory environments.

4.4 Failure Modes and Lesion Responses

The divergent operational mechanics of these two models lead to fundamentally incompatible predictions regarding how cognitive systems degrade when subjected to structural trauma or targeted physical damage. In a modular architecture, the system’s modularity guarantees that structural damage yields crisp, domain-specific deficits. Because modules are computationally encapsulated and functionally autonomous, physical damage can completely sever, destroy, or disable a specific module (e.g., the language production engine, the facial identification processor, or the numerical calculation mechanism) while leaving the surrounding cognitive landscape functioning at pristine, baseline efficiency. The clinical discovery of clean, single and double dissociations in post-stroke patients is celebrated by modular theorists as direct physical verification of this structural prediction.

In contrast, pure connectionist networks, when subjected to damage, exhibit generalized, diffuse failure modes governed by graceful degradation. Because representations are distributed across shared weight matrices, removing a cluster of nodes or severing a portion of connection weights typically produces a soft, proportional loss of precision across all tasks that share those underlying representations, rather than the total, catastrophic abolition of a single, isolated domain. Connectionists, such as David Plaut and Martha Farah, have conducted extensive computational simulations demonstrating that even seemingly clean double dissociations can emerge naturally in damaged, fully distributed connectionist networks without localized modules. Because non-linear mapping surfaces within continuous vector spaces often feature regions of varying computational difficulty or differing topological sensitivity, uniform or random damage to a connectionist network can manifest as an apparent behavioral dissociation. Consequently, connectionists challenge the modular dogma that a double dissociation in neuropsychology provides incontrovertible proof of an underlying, encapsulated mental module.

5. Similarities, Convergences, and Theoretical Overlap

5.1 Shared Mechanistic and Computational Foundations

Despite their fierce, decades-long academic warfare, the Modular and Connectionist paradigms share profound metaphysical, computational, and epistemic commitments that unite them against older, non-scientific philosophies of mind. Both frameworks stand as uncompromising, physicalist repudiations of Cartesian substance dualism. Neither paradigm invokes an immaterial soul, an uncaused will, or a mystical homunculus sitting inside the pineal gland. Both modularists and connectionists agree that the mind is entirely a physical phenomenon, realized within the biophysical structures of the nervous system, and that all mental operations can, in principle, be fully explained through naturalistic, scientific laws.

Moreover, both paradigms are fundamentally computational. They are unified by the core premise of cognitive science: that the mind is an information-processing system whose primary task is to map environmental inputs into behavioral, cognitive, and motor outputs through systematic mathematical or algorithmic transformations. Both paradigms adhere to the methodological imperative of computational modeling. Both reject mere descriptive introspection in favor of rigorous, empirical hypothesis testing, demanding that theoretical assertions be operationalized in mathematically explicit, executable computer simulations and verified through rigorous psychophysical experimentation, behavioral timing studies, and cognitive neuroscience methodologies. In the grand philosophical landscape, the debate between modularity and connectionism is not an existential conflict between science and mysticism; it is an internal, familial dispute over which computational formalism accurately describes the mechanics of human cognition.

5.2 Emergent Modularity in Connectionist Networks

As the debate matured through the 1990s and into the twenty-first century, the stark ideological dichotomy between rigid modularity and undifferentiated connectionism began to soften, revealing rich points of theoretical convergence. One of the most critical insights to emerge from advanced connectionist research is that modularity can be an emergent outcome of connectionist processing rather than an innate architectural constraint. A connectionist network does not have to remain an undifferentiated, homogeneous soup of units; under the right conditions, connectionist architectures naturally differentiate into functionally specialized sub-networks through the normal operation of continuous learning algorithms.

A classic computational illustration of this phenomenon is the “Mixture of Experts” architecture pioneered by Robert Jacobs, Michael I. Jordan, and Geoffrey Hinton (1991). In these architectures, multiple connectionist sub-networks compete for the right to process different subsets of the input data under the coordination of a gating network. As the system trains via gradient descent, the gating network learns to allocate specific classes of inputs to specific sub-networks. The individual networks progressively specialize: one network becomes an expert in processing spatial coordinates, while another specializes in fine-grained visual classification. Crucially, this functional specialization is not pre-programmed or structurally imposed by hand; it emerges spontaneously from competitive learning dynamics and the statistical distribution of the training data. Similarly, continuous self-organizing maps (Kohonen networks) naturally form segregated, localized functional regions that mirror the tonotopic and retinotopic maps found across biological sensory cortices. Thus, connectionism demonstrated that observing modular organization at the macroscopic level does not disprove connectionist mechanics; rather, it shows that connectionism provides the underlying computational machinery through which modularity emerges.

5.3 Hybrid Cognitive Architectures

Recognizing that each paradigm possesses distinct computational strengths and glaring limitations, an increasing number of cognitive theorists and AI researchers have abandoned dogmatic purity in favor of hybrid cognitive architectures. Pure modular-symbolic systems excel at formal mathematical reasoning, recursive syntactic parsing, explicit logical deduction, and robust generalization from single examples via formal rules. However, they are brittle, computationally rigid, and notoriously poor at handling noisy, raw sensory data. Pure connectionist networks excel at continuous perceptual classification, intuitive pattern matching, motor control, and handling noisy, ambiguous real-world environments. However, they struggle with explicit multi-step deductive reasoning, structural systematicity, and explainable causal inference.

To synthesize these capabilities, hybrid architectures systematically integrate both models. In contemporary cognitive science, architectures such as ACT-R (Adaptive Control of Thought—Rational), developed by John R. Anderson, operationalize this synthesis. ACT-R features a macroscopic modular structure composed of specialized, encapsulated modules (a visual module, a declarative memory module, a goal module, and a procedural motor module) coordinated by a central production system. Yet, within these individual modules, the retrieval, activation dynamics, and associative strength of memories are governed entirely by continuous, subsymbolic, connectionist-style mathematical equations. Similarly, in modern artificial intelligence, the surging field of neuro-symbolic AI actively merges deep connectionist neural networks with structured symbolic-modular reasoning engines. In these systems, deep connectionist front-ends process messy, high-dimensional sensory data (such as video feeds or natural language audio) and distill them into discrete, symbolic representations, which are subsequently manipulated by formal, modular logic engines to guarantee safety, factual consistency, and rigorous deductive validity.

6. When to Apply Each Model: Practical and Theoretical Guidance

6.1 Ideal Domains for Modular Hypotheses

The Modular Mind model remains the most scientifically productive and analytically powerful framework when researchers are investigating cognitive phenomena that demonstrate high degrees of structural insulation, behavioral automaticity, evolutionary specificity, and clean clinical dissociability. When approaching a target cognitive capacity, researchers should favor a modular hypothesis under several specific conditions:

• Low-Level Sensory and Perceptual Transduction: In early sensory processing, informational encapsulation is a documented physical reality. Phenomena such as early visual edge detection in V1, primary acoustic tonotopic analysis, and phoneme boundary detection operate via fast, mandatory, encapsulated computational conduits that are impervious to central epistemic penetration.
• Evolutionarily Canalized Survival Behaviors: When studying behavioral adaptations that directly resolved recurrent, life-or-death challenges in ancestral environments, modular frameworks offer high explanatory power. Examples include innate snake and spider fear conditioning, maternal-infant bonding heuristics, kin-directed altruistic calculation, and social cheater-detection mechanisms. These systems appear to operate on dedicated heuristics designed to minimize fatal evolutionary errors.
• Clinical and Diagnostic Neuropsychology: When working with patients suffering from acute vascular accidents (strokes), traumatic brain injuries, or specific genetic disorders (such as Williams syndrome or specific language impairment), modular models provide clinician-researchers with an actionable framework for isolating targeted functional deficits and designing highly focused cognitive rehabilitation protocols.

6.2 Ideal Domains for Connectionist Hypotheses

Conversely, the Connectionist Mind framework is the optimal theoretical and experimental tool when researchers are examining cognitive phenomena characterized by continuous plasticity, statistical induction, holistic contextual integration, and complex pattern completion. Researchers should prioritize connectionist hypotheses under the following empirical circumstances:

• Developmental Trajectories and Plastic Learning: When modeling the gradual acquisition of cognitive skills over developmental time—such as infant vocabulary growth, the transition from overregularization to mastery in irregular past-tense morphology, or the acquisition of motor coordination—connectionist learning algorithms provide unmatched computational fidelity to observed empirical learning curves.
• Continuous Perceptual Categorization and Intuitive Judgment: Tasks that require an agent to rapidly map continuous, noisy, real-world inputs into fuzzy conceptual categories (e.g., distinguishing a dog from a cat in poor lighting, identifying a musical genre, or reading messy human handwriting) are naturally modeled as vector settling processes within continuous connectionist state spaces.
• Engineering High-Performance Artificial Intelligence: In applied technological engineering, connectionist architectures (deep artificial neural networks, transformers, convolutional networks) represent the gold standard for computer vision, automated natural language generation, real-time speech transcription, and multi-modal sensory synthesis.

6.3 Methodological Selection Framework for Researchers

To assist cognitive scientists, neuropsychologists, and computational researchers in selecting the appropriate architectural paradigm for their specific empirical investigations, the following decision-making framework outlines the critical evaluative criteria:

1. Analyze the Degree of Epistemic Permeability:
   • Does the cognitive phenomenon operate invariantly regardless of what the subject explicitly knows, believes, or desires? If yes, adopt a Modular framework (focus on encapsulation and domain specificity).
   • Does the phenomenon dynamically shift, alter its output, or integrate nuanced cross-modal data based on holistic, top-down contextual cues? If yes, adopt a Connectionist framework (focus on distributed settling and parallel constraint satisfaction).
2. Evaluate the Nature of the Representational Target:
   • Does the target task require strict, multi-step, compositional deductive logic where a single broken rule invalidates the output (e.g., formal mathematical proof derivation, legal code interpretation)? If yes, prioritize a Symbolic-Modular architecture.
   • Does the task rely on probabilistic, statistical associations extracted from continuous, noisy real-world data (e.g., identifying emotional valence in human facial expressions)? If yes, prioritize a Connectionist network.
3. Determine the Primary Level of Scientific Explanation:
   • Is the research focused on macroscopic, evolutionary, or high-level ecological functions (David Marr’s computational level)? A Modular model often provides the most tractable, intuitive abstractions.
   • Is the research focused on micro-level computational algorithms, biological plausibility, and direct neural implementation (Marr’s algorithmic and implementational levels)? A Connectionist model provides the required biophysical and mathematical granularity.

7. Frequently Asked Questions

7.1 Can a connectionist network exhibit modular behavior without hardcoded boundaries?

Yes. Extensive computational modeling has conclusively shown that connectionist networks can naturally develop functionally segregated, modular-like behaviors without requiring hardcoded, innate computational boundaries. Through competitive learning algorithms, structural bottlenecks, and specialized loss functions, a continuous neural network can partition its internal representational space into distinct sub-networks that specialize in dedicated tasks. For example, in deep convolutional neural networks trained on complex object classification, early layers spontaneously organize into specialized edge detectors, mid-level layers develop dedicated feature detectors (such as textures or curves), and high-level layers converge into specialized, category-specific nodes (such as face-selective or text-selective units). This emergent modularity demonstrates that macroscopic functional segregation can arise organically from microscopic connectionist mechanics, proving that modular behavior in a biological or artificial system does not necessarily imply the presence of innate, hardcoded, Fodorian modules.

7.2 How does Jerry Fodor’s view of modularity differ from evolutionary psychologists’ view?

The critical difference between Jerry Fodor’s classical modularity and the evolutionary psychologists’ Massive Modularity Hypothesis lies in the scope and extent of modular organization across the cognitive hierarchy. Fodor was a minimalist modularist: he maintained that modularity is strictly confined to peripheral input and output systems (sensory perception, early motor planning, and low-level syntactic parsing). Fodor forcefully argued that central cognitive systems—the computational mechanisms responsible for reasoning, scientific induction, analogy, and holistic belief fixation—are fundamentally non-modular, unencapsulated, and isotropic. He believed that central thought requires evaluating the global web of an agent’s beliefs, making it computationally impossible for central cognition to be encapsulated.

In stark contrast, evolutionary psychologists (such as Leda Cosmides, John Tooby, and Dan Sperber) are maximalist modularists. They reject Fodor’s non-modular central processor, asserting that the mind is modular “all the way through.” Under the Massive Modularity Hypothesis, central cognition is not an unencapsulated void; it is an integrated collection of hundreds of specialized, evolved modules dedicated to ancestral survival challenges, such as mate assessment, kin detection, social cheater identification, and predatory hazard appraisal. Evolutionary psychologists argue that an unencapsulated central system would suffer from catastrophic combinatorial explosion and could never have evolved.

7.3 How do modern deep learning architectures relate to the connectionist mind?

Modern deep learning is the direct, direct computational evolution of the classical 1980s connectionist paradigm. The foundational mathematical primitives of deep learning—artificial neurons, continuous activation functions, matrix multiplications representing connection weights, distributed vector representations, and optimization via the backpropagation of error—are identical to those established by the PDP Research Group. What has changed over the intervening decades is scale, architectural depth, and algorithmic refinement. Contemporary deep learning harnesses immense computational power (GPUs and TPUs), vast datasets, and novel architectural innovations (such as self-attention mechanisms in Transformers, residual connections, and normalization layers) to scale connectionist principles to heights unimagined in the twentieth century.

These modern breakthroughs have empirically vindicated many of connectionism’s boldest theoretical claims. Large language models (LLMs) demonstrate that distributed representations trained on raw statistical associations can capture nuanced semantic structures, generate fluent syntax, and exhibit sophisticated reasoning behaviors without requiring hand-engineered symbolic grammars or explicit, hardcoded modular rules. At the same time, the challenges modern deep learning faces—hallucination, opacity, lack of rigorous deductive systematicity, and data inefficiency—continue to mirror the theoretical criticisms that modular and symbolic cognitive scientists leveled against connectionism decades ago.

7.4 Does brain damage definitively prove the modularity of mind?

No, brain damage does not definitively prove the modularity of mind. While the clinical observation of double dissociations following focal brain lesions has historically served as the primary empirical evidence for modular theory, connectionist cognitive scientists have demonstrated that this inference is methodologically vulnerable. Through rigorous computational simulations, researchers like David Plaut and Martha Farah showed that when distributed, non-modular connectionist networks are subjected to localized, random lesions (such as dropping a percentage of weights or units), the networks can produce clean, statistically significant double dissociations in their behavioral outputs.

This occurs because complex, non-linear mappings between inputs and outputs can create computational bottlenecks or uneven topological sensitivity within a shared, distributed weight matrix. A localized lesion may disrupt a critical convergence pathway for one task while sparing an alternative pathway for another, mimicking the clinical presentation of an isolated, damaged module. Therefore, while neuropsychological lesion data are entirely consistent with modular theories, they do not provide decisive, unambiguous proof of modularity. The behavioral deficit may reflect the disruption of an integrated node within an expansive, brain-wide connectionist network rather than the destruction of an autonomous, encapsulated computational box.

8. Conclusion: Synthesizing Structure and Emergence in Cognitive Science

8.1 Summary of Comparative Strengths and Limitations

The intellectual competition between the Modular Mind and the Connectionist Mind has profoundly enriched our understanding of human and artificial intelligence, revealing the unique explanatory strengths and inherent vulnerabilities of both frameworks. The Modular Mind framework provides a rigorous, highly intuitive account of cognitive automaticity, informational insulation, and functional specialization. Its diagnostic criteria—particularly informational encapsulation and domain specificity—neatly explain why our perceptual systems resist cognitive penetration, why optical illusions remain impervious to our rational knowledge, and why localized brain injuries produce clean, selective cognitive deficits. Furthermore, modularity provides evolutionary biology with a compelling framework for understanding how natural selection engineered specialized adaptations to resolve ancestral survival challenges. However, when pushed to absolute extremes, modularity falters: it struggles to provide a computationally tractable explanation of holistic reasoning, creative cross-domain analogy, contextual fluid intelligence, and the dynamic, plastic capacity of the mammalian brain to reorganize itself following severe neurological injury.

The Connectionist Mind framework, on the other hand, captures the messy, continuous, and biologically plausible reality of cognition. By grounding its primitives in distributed representations, continuous vector mathematics, and parallel dynamic processing, connectionism effortlessly accounts for associative memory, spontaneous generalization, pattern completion, and graceful degradation in the face of noise and physical trauma. Its developmental empiricism demonstrates how complex cognitive competencies—from linguistic morphology to fine-grained visual classification—can emerge organically through statistical exposure to an environment without requiring an exhaustive innate genetic blueprint. Yet, pure connectionism faces enduring limitations: it remains notoriously opaque, frequently struggles to explain the strict systematicity and compositional productivity of formal deductive thought, and often requires staggering quantities of training data to learn tasks that human children master through rapid, rule-governed inference.

8.2 The Future of Integrated Cognitive Architecture

As cognitive science advances through its seventh decade, the historical dichotomy between modular structure and connectionist emergence is increasingly viewed not as an irreconcilable conflict, but as a false choice dictated by looking at different levels of analysis. The definitive resolution of the debate lies in a multi-scale, hierarchical synthesis. Modern cognitive neuroscience reveals that the living brain is simultaneously modular and connectionist: it exhibits clear regional functional segregation at macroscopic cortical scales, while simultaneously relying on massively parallel, recurrent, and distributed connectionist dynamics at microscopic synaptic scales.

The future of cognitive architecture—both in understanding biological minds and in engineering artificial general intelligence—belongs to integrated paradigms that synthesize these complementary virtues. In this emerging consensus, connectionist dynamics provide the underlying computational engine that enables adaptive learning, perceptual fluency, and continuous sensory integration, while modular organization emerges as the stabilizing macroscopic architecture that guarantees functional efficiency, causal reasoning, and structured semantic compositionality. By moving past the tribal hostilities of the past, contemporary cognitive scientists and AI architects are constructing a unified science of mind: an architecture where continuous neural emergence and encapsulated structural specialization cooperate to generate the full, breathtaking spectrum of human intelligence.

References

• Anderson, J. R. (2007). How can the human mind occur in the physical universe? Oxford University Press. https://doi.org/10.1093/acprof:oso/9780195324259.001.0001
• Barkow, J. H., Cosmides, L., & Tooby, J. (Eds.). (1992). The adapted mind: Evolutionary psychology and the generation of culture. Oxford University Press.
• Carruthers, P. (2006). The architecture of the mind: Massive modularity and the flexibility of thought. Oxford University Press. https://doi.org/10.1093/0199207085.001.0001
• Elman, J. L., Bates, E. A., Johnson, M. H., Karmiloff-Smith, A., Parisi, D., & Plunkett, K. (1996). Rethinking innateness: A connectionist perspective on development. MIT Press. https://doi.org/10.7551/mitpress/5935.001.0001
• Farah, M. J. (1994). Neuropsychological inference with an interactive brain: A critique of the “locality” assumption. Behavioral and Brain Sciences, 17(1), 43–61. https://doi.org/10.1017/S0140525X0003322X
• Fodor, J. A. (1983). The modularity of mind: An essay on faculty psychology. MIT Press. https://doi.org/10.7551/mitpress/4737.001.0001
• Fodor, J. A., & Pylyshyn, Z. W. (1988). Connectionism and cognitive architecture: A critical analysis. Cognition, 28(1–2), 3–71. https://doi.org/10.1016/0010-0277(88)90031-5
• Hebb, D. O. (1949). The organization of behavior: A neuropsychological theory. John Wiley & Sons.
• Jacobs, R. A., Jordan, M. I., Nowlan, S. J., & Hinton, G. E. (1991). Adaptive mixtures of local experts. Neural Computation, 3(1), 79–87. https://doi.org/10.1162/neco.1991.3.1.79
• Marr, D. (1982). Vision: A computational investigation into the human representation and processing of visual information. W. H. Freeman and Company.
• Plaut, D. C. (1995). Double dissociation without modularity: Evidence from connectionist neuropsychology. Journal of Clinical and Experimental Neuropsychology, 17(2), 291–321. https://doi.org/10.1080/01688639508405124
• Rumelhart, D. E., McClelland, J. L., & the PDP Research Group. (1986). Parallel distributed processing: Explorations in the microstructure of cognition (Vols. 1 & 2). MIT Press. https://doi.org/10.7551/mitpress/5236.001.0001
• Sperber, D. (2001). In defense of massive modularity. In E. Dupoux (Ed.), Language, brain, and cognitive development: Essays in honor of Jacques Mehler (pp. 47–57). MIT Press.