Loading field map
Cognitive science was born from a simple but unsettling question: how can a physical system—a brain, a machine, or a body—produce thought, perception, and action? Since the mid-twentieth century, researchers from psychology, computer science, philosophy, linguistics, and neuroscience have proposed very different answers. The history of the field is not a steady accumulation of knowledge but a series of competing frameworks, each with its own core commitments, methods, and blind spots. Understanding these frameworks and their relationships is the best way to grasp what cognitive science is and where it is heading.
The first systematic attempt to build a science of mind was Cybernetics, which emerged in the late 1940s from the work of Norbert Wiener and others. Cybernetics treated cognition as a matter of feedback and control: an organism or machine senses its environment, compares the current state to a goal, and adjusts its behavior accordingly. This framework introduced concepts like negative feedback, self-regulation, and circular causality, and it inspired early robots and information-processing models. Cybernetics was deeply interdisciplinary, uniting engineers, biologists, and psychologists around a shared vocabulary of control systems. However, its focus on continuous feedback loops left little room for the kind of abstract, rule-governed reasoning that seemed central to human thought.
That gap was filled by Classical Computationalism, which crystallized at the 1956 Dartmouth Conference and became the dominant framework for the next three decades. Classical Computationalism, often called the "physical symbol system hypothesis," holds that cognition is a form of symbol manipulation: the mind stores and processes discrete, language-like representations according to formal rules. This view was inspired by the digital computer and found powerful expression in the work of Allen Newell, Herbert Simon, and Noam Chomsky. Unlike Cybernetics, which emphasized continuous feedback, Classical Computationalism treated cognition as a step-by-step, serial process operating on internal symbols. It offered a clear, testable research program and generated landmark achievements in problem-solving, language parsing, and early artificial intelligence. For a generation, the mind was a computer.
By the early 1980s, Classical Computationalism had become the orthodoxy, but it faced two very different kinds of pressure from within the computational camp. The first came from Cognitive Architectures, a framework launched by John Anderson's ACT* (1983) and Allen Newell's Soar (1987). Cognitive Architects accepted the symbol-processing premise but argued that the mind's structure—its memory systems, production rules, and learning mechanisms—had to be specified in detail to explain complex human performance. They built unified, often production-system-based models that could simulate everything from skill acquisition to problem-solving. Cognitive Architectures did not reject Classical Computationalism; they narrowed and operationalized it, turning a philosophical thesis into a concrete engineering discipline. They remain active today in areas like intelligent tutoring systems and cognitive modeling.
The second pressure was more radical. Connectionism (Parallel Distributed Processing), launched by the 1986 two-volume work of David Rumelhart, James McClelland, and the PDP Research Group, proposed that cognition emerges from the parallel activity of many simple, neuron-like units. Instead of manipulating discrete symbols, connectionist networks learn by adjusting the strengths of connections between units, gradually capturing statistical regularities in the input. This framework directly competed with Classical Computationalism: it denied that the mind needs explicit rules or local symbols, and it offered a more brain-like, graded, and fault-tolerant alternative. The rivalry was intense throughout the late 1980s and 1990s, with debates over whether connectionist networks could handle structured reasoning (e.g., language syntax) or whether they merely implemented symbol processing at a lower level. Connectionism did not replace Classical Computationalism, but it permanently expanded the field's sense of what a cognitive model could look like.
While the computationalists and connectionists argued over the right way to model internal processes, a different kind of rebellion was brewing. Embodied Cognition, which gained momentum in the late 1980s and early 1990s, argued that the classical and connectionist approaches both made the same mistake: they treated cognition as something that happens entirely inside the head. Embodied Cognition insisted that the body—its morphology, sensorimotor capacities, and interactions with the environment—shapes thought in fundamental ways. This was not a minor tweak but a direct reaction against the internalist assumptions of both Classical Computationalism and Connectionism. Researchers like Eleanor Rosch, Francisco Varela, and later Andy Clark showed that perception, memory, and even abstract reasoning are grounded in bodily experience and action.
Enactivism, developed by Varela, Thompson, and Rosch in their 1991 book The Embodied Mind, pushed this critique further. Enactivism holds that cognition is not a matter of representing a pre-given world but of enacting a world through the history of an organism's sensorimotor couplings. Where Embodied Cognition often still uses representations (just more body-based ones), Enactivism is more radical: it rejects the very idea of mental content and argues that meaning arises from the dynamic, structural coupling between organism and environment. This places Enactivism in a deeper disagreement with Classical Computationalism and Connectionism than Embodied Cognition typically is.
Dynamical Cognitive Science, which emerged around 1994 from the work of Esther Thelen, Linda Smith, and others, offered yet another alternative. Drawing on dynamical systems theory, this framework treats cognition as a continuous, self-organizing process that unfolds in real time through the interaction of brain, body, and environment. Dynamical Cognitive Science explicitly rejects the need for internal representations or discrete computational steps, viewing cognition as an emergent property of coupled nonlinear systems. It shares with Embodied Cognition and Enactivism a focus on embodiment and time, but it is distinguished by its mathematical toolkit (differential equations, phase spaces, attractors) and its emphasis on development as a key window into cognitive change. Together, these three frameworks—Embodied Cognition, Enactivism, and Dynamical Cognitive Science—formed a coordinated revolt against the internalist, representation-heavy views that had dominated the field, though they disagreed among themselves about how radical the break needed to be.
At the same time that embodied and dynamical approaches were gaining ground, two other frameworks offered very different ways to move beyond Classical Computationalism. Evolutionary Psychology, launched by Leda Cosmides and John Tooby in the late 1980s and popularized by Steven Pinker in the 1990s, argued that the mind is not a general-purpose symbol processor but a collection of specialized modules shaped by natural selection to solve ancestral problems. This framework competed directly with Classical Computationalism's assumption of domain-general reasoning and with Connectionism's emphasis on learning from scratch. Evolutionary Psychology insisted that many cognitive capacities—for mate selection, cheater detection, or folk physics—are innate, modular, and largely fixed. Its critics charged that it relied on speculative adaptive stories and underestimated cultural and developmental plasticity.
Bayesian Cognitive Science, which took off in the 1990s, offered a different kind of alternative. Instead of positing innate modules, Bayesian approaches treat cognition as a form of probabilistic inference: the mind combines prior knowledge with noisy sensory evidence to compute the most likely interpretation of the world. This framework drew on formal probability theory and quickly became a powerful tool for modeling perception, motor control, language, and even high-level reasoning. Unlike Evolutionary Psychology, Bayesian Cognitive Science did not reject learning; it formalized it. And unlike Classical Computationalism, it did not assume that the mind manipulates discrete symbols; it assumed that the mind represents probability distributions. Bayesian Cognitive Science also influenced Connectionism: starting around 2000, researchers began using Bayesian methods to learn the structure of connectionist networks, creating hybrid models that combined the flexibility of neural networks with the normative rigor of Bayesian inference. This cross-fertilization is one of the clearest examples of frameworks not just competing but actively shaping each other.
While many frameworks focused on high-level cognition or behavior, Theoretical Neuroscience, emerging in the 1990s, aimed to ground cognitive explanations in the biophysics and dynamics of neural circuits. Researchers like Peter Dayan, Larry Abbott, and William Bialek developed mathematical models of neural firing, synaptic plasticity, and population coding. Theoretical Neuroscience did not compete directly with Classical Computationalism or Connectionism; instead, it provided a lower-level infrastructure that could, in principle, support or constrain higher-level theories. It coexists with other frameworks by offering mechanistic accounts of how neural activity implements cognitive functions.
Predictive Processing, which began to crystallize around 1999 with work by Rajesh Rao, Dana Ballard, and later Karl Friston, represents a major synthesis. Predictive Processing argues that the brain is fundamentally a prediction engine: it constantly generates top-down predictions about sensory input and updates its internal models based on prediction errors. This framework integrates Bayesian Cognitive Science's emphasis on probabilistic inference with Theoretical Neuroscience's focus on neural implementation. It also resonates with Embodied Cognition and Enactivism by treating perception and action as two sides of the same coin (action minimizes prediction error by sampling the world). Predictive Processing has become one of the most active and unifying frameworks in contemporary cognitive science, offering a single principle—prediction error minimization—that can explain perception, action, attention, learning, and even consciousness.
The most recent major framework, Deep Learning, exploded onto the scene in 2012 when deep neural networks achieved dramatic breakthroughs in image recognition, speech processing, and game playing. Deep Learning is a direct descendant of Connectionism: it uses multi-layer neural networks trained with backpropagation, but at a vastly larger scale, with more data, and with new architectural innovations (convolutional layers, recurrent networks, transformers). It has transformed artificial intelligence and, in turn, cognitive science. Deep Learning does not reject the insights of earlier frameworks; it revives and extends connectionist ideas while also providing powerful tools for testing Bayesian and predictive processing models. However, it also raises new questions: Do deep networks learn in human-like ways? Do they need symbolic scaffolding? Can they be integrated with embodied and dynamical approaches? Deep Learning is currently a leading framework not because it has solved cognition but because it offers unprecedented empirical success and a flexible modeling platform.
Today, cognitive science is a pluralistic field. The leading frameworks—Predictive Processing, Deep Learning, Bayesian Cognitive Science, and Embodied Cognition—each have distinct strengths. Predictive Processing offers a unifying theoretical principle that spans perception, action, and learning. Deep Learning provides powerful, scalable models that achieve human-level performance on many tasks. Bayesian Cognitive Science supplies normative, mathematically rigorous accounts of inference and learning. Embodied Cognition reminds the field that cognition is situated, active, and shaped by the body. These frameworks often overlap: Bayesian methods are used to analyze deep networks; predictive processing models are implemented in neural networks; embodied cognition researchers draw on dynamical systems theory.
Yet deep disagreements remain. The most persistent tension is over the role of internal representations. Classical Computationalism, Cognitive Architectures, and Bayesian Cognitive Science all assume that the mind represents the world in some form. Embodied Cognition, Enactivism, and Dynamical Cognitive Science question whether representation is necessary or even useful. A second tension concerns domain-specificity: Evolutionary Psychology argues for innate modules, while Connectionism and Deep Learning emphasize domain-general learning mechanisms. A third tension is about the right level of explanation: Theoretical Neuroscience and Predictive Processing aim for neural mechanisms, while Cognitive Architectures and Bayesian models often stay at a more abstract, algorithmic level. These disagreements are not signs of weakness; they are the engine of the field. Cognitive science advances not by converging on a single framework but by forcing each framework to confront the phenomena that the others handle best.