Loading field map
Cognitive neuroscience has always been pulled between two opposing intuitions. One is that different mental functions—language, memory, vision—are carried out by dedicated brain regions. The other is that the brain works as a whole, with cognition emerging from interactions across many areas. This tension between localization and distributed processing has driven the field since its origins, and each new explanatory framework has been, in part, a response to the limits of its predecessors.
The first systematic attempt to map mental functions onto brain tissue was Localizationism, which took shape in the early 1800s. Paul Broca’s 1861 discovery that damage to a specific left frontal region caused speech production deficits, and Carl Wernicke’s later identification of a temporal area for speech comprehension, seemed to confirm that cognitive functions occupy fixed cortical addresses. Localizationism remains active today in the widespread use of functional neuroimaging to identify regionally specific activations for perception, attention, memory, and language.
Yet even in its heyday, localizationism faced a powerful challenge. In the 1920s and 1930s, Karl Lashley’s experiments on rats—removing varying amounts of cortex before training them on mazes—led him to propose Holism and Equipotentiality. Lashley argued that cortical tissue was largely interchangeable: the size of a lesion mattered more than its location, and any intact region could take over lost functions. Holism directly rejected the localizationist assumption that each cognitive operation has a fixed neural address. Although the equipotentiality claim was later qualified—it worked best for simple learning tasks—the holist insistence on distributed function never entirely disappeared. It resurfaced decades later in connectionist and network models.
A more nuanced alternative to pure localization emerged in 1874, when Wernicke proposed Disconnectionism. He argued that cognitive functions depend not only on specialized cortical centers but also on the white-matter pathways that connect them. A language deficit, for example, could arise from a lesion in a center or from a break in the fiber tract linking two centers. Disconnectionism preserved the localizationist commitment to specialized regions while adding a crucial relational dimension: cognition requires coordination across regions. This framework proved remarkably durable and is still invoked in modern studies of white-matter tracts and their role in neurological syndromes.
A century later, Cognitive Neuropsychology (1970–present) gave disconnectionist logic a new, more abstract form. Instead of focusing on anatomical pathways, cognitive neuropsychologists inferred the architecture of normal cognition from patterns of spared and impaired performance in brain-damaged patients. The key method was the double dissociation: if patient A can do task X but not Y, while patient B shows the opposite pattern, the two tasks must rely on at least partially separable cognitive components. This approach allowed researchers to build functional models of reading, memory, and attention without needing to know exactly where in the brain those components were located. Cognitive neuropsychology thus coexisted with localizationist neuroimaging, each addressing a different level of analysis—functional architecture versus neural implementation.
The 1980s brought a decisive shift: cognitive neuroscience became computational. But the computational turn was not a single movement; it split into two frameworks with different assumptions about how the brain computes.
Computational Cognitive Neuroscience (1982–present) grew out of David Marr’s influential tri-level analysis, which distinguished computational goals, algorithmic representations, and physical implementation. Researchers in this tradition build mechanistic models of specific cognitive processes—such as visual object recognition or working memory—using biologically plausible neural networks whose parameters are tuned to fit behavioral and neural data. The framework is algorithm-focused: it asks what computations a brain region performs and how those computations are realized by neural circuits.
At nearly the same time, Connectionism (1986–present) emerged from the Parallel Distributed Processing (PDP) movement led by David Rumelhart, James McClelland, and their colleagues. Connectionist models are also neural networks, but they emphasize learning from experience rather than pre-specified algorithms. A typical connectionist network starts with random connection weights and gradually adjusts them through error-driven learning (backpropagation) to perform a task. The resulting representations are distributed across many units and are not localized to any single node. Connectionism thus revived the holist intuition that knowledge is spread across many processing elements, and it directly challenged the symbolic, rule-based models that dominated cognitive science at the time.
Both computational frameworks remain active, but they have diverged in practice. Computational cognitive neuroscience tends to build more anatomically detailed models and to test them against neuroimaging and electrophysiological data. Connectionism, especially in its modern deep-learning form, focuses on large-scale learning and has become a dominant tool in artificial intelligence. The two traditions overlap in their use of neural networks but differ in their primary explanatory target: mechanistic understanding versus learning dynamics.
By the 1990s, a third computational approach began to take shape. Network Neuroscience (1990–present) treats the brain as a complex network of nodes (neurons or brain regions) and edges (structural or functional connections). Using graph theory, network neuroscientists measure properties such as modularity, hub centrality, and small-world architecture. This framework competes with classical localizationism by arguing that cognitive functions are not properties of individual regions but emergent properties of network topology. A region’s role depends on its pattern of connections, not just its local cytoarchitecture. Network neuroscience also absorbs the disconnectionist insight that disrupted connectivity can cause cognitive deficits, but it generalizes that idea to whole-brain network dynamics.
Network neuroscience has become a major force in the field, especially for studying large-scale brain organization. It provides a vocabulary for describing how the brain balances integration and segregation—a question that earlier frameworks could only address in qualitative terms.
The most recent frameworks push beyond the classical computational model in two different directions.
Predictive Processing (1999–present) proposes that the brain is fundamentally a prediction engine. It continuously generates top-down predictions about sensory input and updates its internal models based on prediction errors. This framework, rooted in Bayesian inference and hierarchical generative models, competes directly with Computational Cognitive Neuroscience’s emphasis on feedforward, algorithm-driven processing. In predictive processing, perception, action, and attention are all unified under a single principle: minimizing free energy or prediction error. The framework has been applied to everything from visual illusions to psychosis, and it challenges the modularity assumption that perception and cognition are separate systems. Predictive processing remains a lively research program, though its claims about neural implementation are still debated.
Embodied Cognitive Neuroscience (2000–present) takes a different kind of departure. It argues that cognition cannot be understood by studying the brain in isolation; the body and the environment are constitutive parts of the cognitive system. Perception is shaped by action possibilities, memory is tied to sensorimotor experience, and abstract concepts are grounded in bodily interactions. This framework reacts against the brain-centric assumptions of both localizationism and computational cognitive neuroscience, which treat cognition as an internal, representational process. Embodied cognitive neuroscience draws on behavioral experiments, neuroimaging, and robotics to show that neural activity is modulated by bodily states and ecological context. It remains a minority perspective within cognitive neuroscience, but it has forced the field to take seriously the idea that the brain is not the whole story.
Today, no single framework dominates cognitive neuroscience. Localizationism persists in the routine use of fMRI to map function to region. Disconnectionism lives on in diffusion-tractography studies of white-matter pathways. Cognitive neuropsychology continues to inform clinical assessment and theoretical models of cognitive architecture. Computational cognitive neuroscience and connectionism provide complementary modeling tools. Network neuroscience offers a systems-level language that integrates many of these traditions. Predictive processing and embodied cognition represent active frontiers that challenge settled assumptions.
The leading frameworks agree on at least one point: the brain is a complex, interactive system, and understanding cognition requires linking multiple levels of analysis—from molecules to networks to behavior. Where they disagree is on what the fundamental explanatory unit should be. Is it a region (localizationism), a pathway (disconnectionism), a distributed pattern (connectionism), a network topology (network neuroscience), a generative model (predictive processing), or a body-environment system (embodied cognitive neuroscience)? These are not just methodological preferences; they reflect deep differences about what cognition is and how it should be studied. The history of cognitive neuroscience suggests that the most productive research often combines insights from competing frameworks rather than choosing one over the others.