RESEARCH DOSSIER
Connectionist and Neural Network Models of Cognition
Abstract. Connectionist and neural network models investigate how cognition can be explained through distributed representations, learning-dependent connectivity, and interacting computational processes. This monograph critically reconstructs the development of perceptron theory, interactive activation, parallel distributed processing, recurrent learning, complementary memory systems, and contemporary deep networks. The evidence review examines primary computational studies and experimental findings in visual and lexical perception, reading, morphological generalization, semantic cognition, memory interference, cortical object recognition, and language processing. Methods include experimental manipulation, formal model comparison, representational similarity analysis, out-of-distribution tests, and cross-level neural and behavioral validation. The synthesis finds that explicit network models make substantial contributions to cognitive explanation when they generate distinctive predictions beyond trained datasets; important open questions concern compositional transfer, learning histories, mechanistic identifiability, biological plausibility, and robustness across people and tasks.
Keywords: connectionism; neural network models of cognition; parallel distributed processing; distributed representations; interactive activation; cognitive modeling; backpropagation; recurrent neural networks; semantic cognition; reading models; compositional generalization; complementary learning systems; representational similarity analysis; human–AI cognitive comparison
Scientific Object and Explanatory Criteria
Connectionist and neural network models of cognition explain psychological functions through the coordinated activity of interconnected computational units whose states and connection strengths determine how information is transformed. Their scientific object is a set of hypotheses concerning representation, learning, processing dynamics, generalization, and the relationship between experience and behavioral performance. An adequate model specifies inputs, internal transformations, outputs, training experience, and the observable phenomena it purports to explain. A network that classifies images accurately is evidence of engineering performance; a network that predicts human errors, response times, and neural response structure provides evidence relevant to cognitive explanation Yang & Wang (2020) Saxe et al. (2021).
The central methodological question is whether a particular network architecture provides an informative mechanistic account of a particular cognitive phenomenon. Answering it requires distinguishing computational goals from algorithms and biological implementation. Models may be descriptive approximations of behavioral regularities, learning theories, candidate process mechanisms, or explicit hypotheses about cortical circuitry. The same architecture may serve different explanatory purposes under different experimental designs. Evaluation should therefore identify which level of explanation the investigators actually test, the alternative models under comparison, and the data withheld from model construction Marr (1982) Wilson & Collins (2019).
A research dossier devoted to connectionism has a narrower remit than the general science of mathematical psychology. It examines the explanatory commitments of networks whose internal units jointly represent information and whose processing or learning follows defined connectionist principles. Bayesian approaches, symbolic cognitive architectures, and dynamical systems enter the analysis when they provide rival explanations or compatible formal tools. Likewise, the generic use of machine learning to predict psychological outcomes is considered only when the learning model makes claims about cognitive processes, developmental trajectories, or task-generalization mechanisms Fodor & Pylyshyn (1988) Lake et al. (2017).
The evidence assembled here consists of landmark computational investigations, behavioral experiments used to constrain models, analyses of neuropsychological disorders, and contemporary cross-level comparisons of machine representations with neural and human behavioral measurements. It is a critical integrative research monograph rather than a preregistered systematic review. The synthesis emphasizes source-specific findings and separates experimental observations, simulation results, theoretical deductions, and hypotheses that await decisive experiments. Such differentiation is indispensable because the apparent explanatory breadth of neural networks can conceal considerable flexibility in training objectives, stimuli, parameterization, and readout procedures Pitt & Myung (2002) Schrimpf et al. (2020).
Historical Origins of Network Thought
The historical emergence of connectionism joins traditions in mathematical psychology, neurophysiology, cybernetics, and computational learning. Hebbian proposals linked experience-dependent changes in effective connectivity to the persistence of learned activity patterns; formal neuronlike models subsequently made it possible to investigate the collective consequences of simple local operations. These early claims were theoretical rather than measurements of individual synaptic changes during human reasoning. Their enduring contribution was to turn learning and psychological organization into questions about the relation between connection structure, changing activation, and observable performance Rosenblatt (1958) Hopfield (1982).
Frank Rosenblatt’s 1958 perceptron paper asked how information could be stored and organized by a hypothetical nervous system and introduced an explicit statistical and computational model. It showed that a learning device could be investigated mathematically as a candidate explanation of perceptual recognition and generalization. The original proposal included assumptions about connectivity, statistical organization, and adaptation that should not be collapsed into a modern marketing description of a single-layer classifier. Its scientific importance lies in introducing an operational research program linking learning rules to observable behavior Rosenblatt (1958).
In 1981 McClelland and Rumelhart developed an interactive activation model of the word-superiority phenomenon, in which visual features, letters, and word representations exert excitatory and inhibitory influences across levels. Hopfield’s 1982 analysis of collective network dynamics showed how distributed interactions could support content-addressable recall from incomplete cues. These developments broadened the relevant computational questions from simple classification toward contextual perception, attractor dynamics, and memory retrieval. Both families made explicitly inspectable predictions, although their state variables and mathematical assumptions differed substantially McClelland & Rumelhart (1981) Hopfield (1982).
The publication of Parallel Distributed Processing in 1986 consolidated a program in which cognition emerged from interactions among many elementary processing units and knowledge was encoded in connection strengths. In the same year Rumelhart, Hinton, and Williams described error backpropagation as a procedure for learning useful hidden representations. These works have to be understood as historical sources of specific mechanisms, not as proof that all subsequent deep-learning architectures implement human cognition. The decisive transition was methodological: representations could be learned from structured training experience and tested against patterns of psychological performance Rumelhart et al. (1986) Rumelhart & McClelland (1986).
Perceptrons, Associative Memory, and Parallel Distributed Processing
A perceptron establishes a mapping between an input vector and a response by combining weighted signals and applying an activation or decision rule. In modern notation a simple unit can be expressed as y = f(w·x + b), where x is the input vector, w the learned weights, b a bias, and f the activation function. This formula captures a computational operation, not a literal description of a biological neuron. The relevance of the model to psychological science depends on the variables selected as inputs and outputs and on the extent to which learning predicts human discrimination and transfer behavior Rosenblatt (1958) Yang & Wang (2020).
Associative or attractor networks replace a purely feedforward mapping with a dynamical process in which activity evolves under recurrent connections. Hopfield demonstrated that stored patterns can behave as stable states of a system: a partial input may lead the network toward a complete stored configuration. This supplies a precise account of pattern completion at the level of model dynamics. It does not follow that a particular hippocampal or cortical memory operates according to the original mathematical energy function; such an implementation claim requires separate neural measurements, temporal constraints, and perturbational tests Hopfield (1982) McClelland et al. (1995).
The parallel distributed processing tradition adds multi-layer processing, graded activation, and representations shared across multiple items or experiences. Because a single weight participates in many operations, learning one item can alter the processing of other items. This shared-parameter organization predicts both beneficial transfer and interference. The crucial empirical issue becomes which relations among stimuli are encoded by the network and whether the resulting transfer resembles human performance in unfamiliar conditions. These claims can be examined in perceptual, lexical, semantic, and memory tasks without assuming that a network is a transparent model of the entire mind Rumelhart & McClelland (1986) Seidenberg & McClelland (1989).
Connectionism should be treated as a family of scientific models rather than a single architecture or doctrine. Localist interactive networks, distributed feedforward networks, recurrent dynamical networks, complementary memory systems, convolutional visual models, and contemporary language networks have different internal commitments. Their shared explanatory theme concerns interactions among units and learning-dependent organization. Whether a particular model learns through biologically realistic local signals, globally propagated errors, unsupervised objectives, or externally supplied labels is an additional substantive question. Theoretical comparisons must preserve these differences instead of attributing the properties of one implementation to all connectionist models Rumelhart et al. (1986) Yang & Wang (2020).
Mathematical Anatomy of a Cognitive Network
A neural network model is specified by a computational graph, parameters, activation dynamics, a learning rule, and a mapping from internal activity to psychological measurements. For a layered feedforward system, activations can be written h(l) = f(W(l)h(l−1) + b(l)). Each weight matrix determines how features at one processing stage influence the next. Psychologically meaningful interpretation requires specifying whether the input codes are raw images, orthographic features, phonological segments, semantic properties, or experimental variables. Two networks with identical graph structures can support different cognitive explanations when trained on different environments Yang & Wang (2020) Rumelhart & McClelland (1986).
Learning typically minimizes a discrepancy between model outputs and targets, or maximizes an objective defined over experience. Backpropagation computes how changes in weights affect the selected loss and thereby enables gradient-based optimization of hidden representations. Rumelhart, Hinton, and Williams established that this procedure can produce internal features useful for the task without hand-specifying every feature. Yet the computational efficiency of gradient learning is logically separate from its biological plausibility. An explanatory model must disclose its supervision, update signals, training schedule, and assumptions about information available to the learner Rumelhart et al. (1986) Saxe et al. (2021).
Recurrent models introduce feedback over time, making internal states depend on earlier states as well as current inputs. Such recurrence supports sustained context, sequence sensitivity, and attractor phenomena; it can also create unstable dynamics, long-term dependencies, or training difficulties. Elman’s simple recurrent networks used contextual feedback to show that structured temporal information can be represented in distributed hidden activity. A cognitive prediction therefore depends on the actual sequence distribution, temporal granularity, and readout rule. Demonstrating a fit to an average response curve is insufficient to establish the proposed internal dynamics Elman (1990) Pitt & Myung (2002).
Model complexity comprises more than the number of weights. Architecture, inductive biases, regularization, training data, augmentation, optimization, and output decoding all affect possible behavior. Flexible networks can interpolate training samples while failing on experimental transformations that humans handle routinely. Conversely, a constrained smaller network may generate more informative falsifiable predictions. The scientific task is to identify parameters and mechanisms that remain stable across independently acquired data and experimental interventions. Held-out prediction, model recovery, and preregistered contrasts can prevent apparently impressive fits from being confused with explanatory identification Wilson & Collins (2019) Pitt & Myung (2002).
Representation: Localist, Distributed, and Sparse Codes
A localist representation assigns a unit or small set of units a relatively explicit correspondence with a recognizable feature, letter, word, or concept. The interactive activation model of letter perception exemplifies the interpretive advantages of recognizable representational levels: activation and inhibition can be followed across visual features, letter candidates, and word candidates. Such transparency is purchased through representational assumptions imposed by the investigator. The model’s explanatory merit lies in the resulting predictions about contextual facilitation and competition, rather than in the proposition that each designated unit identifies a single biological cell McClelland & Rumelhart (1981).
A distributed representation encodes an item through the pattern of activation across many units, with individual units contributing to representations of multiple items. Similar items can share subpatterns, enabling graded similarity and generalization. In semantic cognition, this approach permits knowledge of objects to reflect overlapping relations among perceptual and functional features without requiring an isolated symbolic entry for every fact. The structure of the learned representational space is a testable hypothesis: similarity judgments, property-verification errors, priming patterns, and neuropsychological dissociations can all constrain it Rogers et al. (2004) Rogers & McClelland (2004).
Sparse representations occupy an intermediate position: only a limited subset of units is active for a given event, while many overlapping subsets are possible across events. Sparse coding, pattern separation, and distributed overlap are distinct dimensions and should be explicitly operationalized. The complementary learning systems account uses differences in coding and learning rates to address rapid episodic encoding and gradual extraction of statistical structure. These proposals are illuminated by patterns of selective memory impairment but are not proved by the fit of a single simulated lesion. Models require comparison across multiple memory tasks and temporal scales McClelland et al. (1995).
Representations inferred from trained networks can be transformed without changing all observable predictions. A unit-by-unit interpretation may therefore depend on arbitrary basis choices, regularization, and readout constraints. Representational similarity analysis provides a more abstract way to compare relationships among stimuli across models and brains. However, shared similarity structure alone cannot determine a unique internal computation. A strong explanatory claim triangulates geometry with specific learning histories, response distributions, temporal activity, and causal interventions. Interpretability is a graded empirical achievement, not an intrinsic guarantee supplied by the word distributed Kriegeskorte et al. (2008) Hebart et al. (2020).
Learning Rules, Optimization, and Development
Connectionist models offer an explicit account of how performance changes with experience. The modeler chooses a curriculum or exposure distribution and an update process; performance then evolves as weights change. This makes learning trajectories and generalization gradients available for experimental comparison. Seidenberg and McClelland’s reading network illustrated how a single trained system could exhibit changes relevant to early and skilled visual word recognition. The correspondence of a model’s training cycles with developmental age, however, is an additional empirical assertion requiring evidence about exposure, experience, and prior skills Seidenberg & McClelland (1989).
Learning objectives matter because they specify what counts as success for the computational system. Supervised error correction, predictive sequence learning, self-supervised reconstruction, and reinforcement feedback lead to different internal organizations even when networks have similar numbers of units. Cognitive theories must ask whether the learning signal is available in the child’s or adult’s environment and whether its informational demands are realistic. A model trained on extensive labeled data cannot be taken as evidence for rapid human category acquisition without a matched comparison of sample efficiency and inductive structure Rumelhart et al. (1986) Lake et al. (2017).
The mathematical analysis of deep linear networks by Saxe, McClelland, and Ganguli demonstrated how staged acquisition of semantic distinctions can emerge from learning dynamics and the statistical structure of inputs. Its theoretical value is the derivation of conditions that produce representational transitions rather than mere post hoc resemblance to a curve. The analysis uses explicitly simplified architectures, which clarifies what can and cannot be generalized to nonlinear biological networks. Such work helps convert the vague statement that networks develop concepts into formal hypotheses about experience, timing, and representational geometry Saxe et al. (2019).
Developmental claims are strongest when models make predictions about individual variation and changes across controlled stages of training. Researchers can compare trajectories of errors, item-frequency effects, sensitivity to exceptions, and generalization to untrained items. A network may fit adult endpoint performance while producing an implausible order of acquisition, or it may mimic early-stage errors for the wrong reason. Evidence must therefore constrain learning path, final competence, and transfer jointly. The modeling literature demonstrates the usefulness of this strategy while exposing persistent ambiguities about environmental input and the correspondence between simulation epochs and lived learning Seidenberg & McClelland (1989) Saxe et al. (2019).
Interactive Activation and Contextual Perception
The word-superiority phenomenon illustrates why perceptual processing became a central test case for connectionism. When identifying a letter, observers can benefit from the letter appearing in a meaningful or pronounceable context. An interactive activation account proposes that lower-level visual evidence and higher-level word constraints jointly influence the activation of competing letter hypotheses. McClelland and Rumelhart’s 1981 model represented separate feature, letter, and word levels and made the pattern of contextual effects a target of computer simulation. This was an explicit explanation of an experimental effect rather than a general claim that context always improves accuracy McClelland & Rumelhart (1981).
The mechanism depends on graded activation, competition, and feedback. Compatible visual features excite letter units, competing alternatives inhibit one another, and word-level evidence can affect letter-level availability. The psychological interpretation is conditional on the design of the letter-report task, stimulus timing, and the distinction between real words and pronounceable nonwords. In this tradition, an experimentally manipulated context has a diagnostic role because it can reveal the relative contribution of bottom-up signals and learned contextual constraints. The model’s specific connections are commitments that can be challenged with novel display conditions McClelland & Rumelhart (1981).
An important interpretive limit concerns the relationship between simulation and perception itself. A model may account for a behavioral advantage through activation dynamics, decision thresholds, or response selection, each with different implications for the locus of contextual influence. Measurements made after a brief display do not automatically identify the exact neuronal stage at which context operates. The relevant evidence therefore includes time-sensitive experimental designs, response distributions, and alternative accounts of decisional bias. This is an example of why model comparison must be organized around discriminating predictions rather than verbal similarity Pitt & Myung (2002) Wilson & Collins (2019).
Contemporary hierarchical vision networks revisit the same general explanatory issue at a different scale. Their representations may acquire contextual and category structure through large-scale training, but this changes the empirical questions: can the model predict individual stimulus confusions, timing effects, and effects of controlled contextual interventions? A model trained for object recognition is not automatically a theory of letter perception. Connectionist history is most scientifically productive when individual paradigms remain distinct and the tests that supported or constrained early mechanisms are preserved during architectural modernization Yamins et al. (2014) Kriegeskorte (2015).
Visual Word Recognition and Reading Aloud
English reading is a quasi-regular mapping problem because spelling–sound correspondences contain both recurring regularities and exceptions. Seidenberg and McClelland developed a distributed developmental network connecting orthographic and phonological representations through learned hidden units. Their 1989 article investigated whether a common learning procedure could explain variation with lexical properties and changes accompanying learning, including pronunciation of unfamiliar strings. The evidential contribution lies in formalizing the role of item frequency, consistency, and shared statistical structure in reading, not in demonstrating that all reading behavior is reducible to one mapping Seidenberg & McClelland (1989).
Plaut, McClelland, Seidenberg, and Patterson’s 1996 investigation refined the account of skilled and impaired reading in quasi-regular domains. Networks trained with structured representations processed regular and exceptional words while also pronouncing pronounceable nonwords. The authors examined frequency and consistency effects and simulated patterns relevant to acquired surface dyslexia. These results indicate that graded shared constraints can account for phenomena sometimes interpreted as evidence for discrete lexical and sublexical routes. The results remained simulations tied to specified training conditions and neuropsychological interpretations Plaut et al. (1996).
Competing computational models remain essential to assessing this claim. Coltheart and colleagues presented the dual-route cascaded model in 2001, implementing separate procedures within a comprehensive model of visual word recognition and reading aloud. It was evaluated against lexical decision, naming latencies, and other behavioral effects, and offered alternative explanations of several influential results. Comparing the two approaches requires matching tasks and parameter constraints and distinguishing successful reproduction of established effects from independently predicted data. The existence of a successful distributed model therefore does not, by itself, eliminate structured route-based accounts Coltheart et al. (2001) Plaut et al. (1996).
Neuropsychological evidence sharpens this comparison because distinct acquired impairments can selectively affect exception reading, nonword reading, semantic support, and other processes. A lesion simulation is informative only if the type and location of damage are justified independently and the model predicts a constellation of deficits and spared functions. Flexible damage parameters can produce superficially similar syndromes across theories. Critical tests include graded severity, item-level error profiles, recovery and generalization, and consistency with longitudinal observations. Connectionist reading models remain a major example of fruitful dialogue between learning simulation, experimental psycholinguistics, and neuropsychology Plaut et al. (1996) Rogers et al. (2004).
Language Inflection, Rules, and Systematic Generalization
The English past-tense debate turned regular and exceptional morphology into an explicit test of learning theory. Connectionist accounts explored whether mappings between phonological properties of verb stems and past-tense forms could acquire rulelike performance and developmental error patterns from exposure. Pinker and Prince’s 1988 analysis challenged the adequacy of a particular parallel distributed processing treatment and argued that its predictions depended on representational and training choices. The scientific significance of the controversy lies in forcing researchers to specify the kind of generalization they intend to explain Pinker & Prince (1988).
The decisive distinctions concern memorized similarities, productive morphological classes, and operations that apply to arbitrary new instances. A model may correctly classify many trained regular verbs yet fail on a new form outside the neighborhood represented in its experience. Conversely, strong performance on carefully constructed novel cases is evidence that the model learned a generalizable transformation under the tested conditions. Neither observation alone identifies a mentally represented symbolic rule, because several algorithms can implement overlapping input–output relations. The relevant evidence consists of structured generalization sets, developmental profiles, and processing effects Pinker & Prince (1988) Marcus (1998).
Fodor and Pylyshyn’s 1988 critique argued that classical cognitive architecture addresses compositional structure and systematicity in ways that must be explained by connectionist alternatives. Marcus’s 1998 critique further targeted the capacity of certain eliminative connectionist architectures to extend universally quantified relations to arbitrary novel elements. These works are arguments about explanatory adequacy under specified formal assumptions; they should not be interpreted as demonstrations that networks cannot learn any compositional structure. Subsequent comparisons need to isolate which representational resources, binding operations, and training regimes a given system actually provides Fodor & Pylyshyn (1988) Marcus (1998).
For psychology, the unresolved issue is not a contest between fashionable labels but the explanation of human productivity. Controlled tests can use invented words, relation reversals, syntactic role substitutions, or new combinations of familiar constituents to dissociate associative similarity from systematic manipulation. Training-set coverage, cue frequency, and task instructions must be documented because they change the demand on the model. The most informative comparison examines error distributions and transfer gradients in humans and multiple classes of trained systems, including structured and hybrid models. Such experimental work connects connectionism to contemporary theories of conceptual and linguistic compositionality Lake et al. (2017) Pinker & Prince (1988).
Recurrent Networks, Temporal Context, and Prediction
Cognition unfolds through time, and any model of language, action, or memory must explain how earlier information influences later processing. Elman’s 1990 recurrent network study introduced context units that preserved a representation of earlier hidden-layer activity, enabling the system to use temporally extended dependencies. Simulations examined problems from simple temporal mappings to syntactic and semantic structure in word sequences. The contribution was a concrete demonstration that context-sensitive internal representations could arise through sequence learning. It did not prove that human grammar is acquired solely through the particular recurrent architecture Elman (1990).
A recurrent state can be described as h(t) = f(Wx(t) + Uh(t−1) + b), with the new state incorporating input and previous context. This formulation permits history-dependent responses without storing every earlier item in a separate symbolic buffer. The available memory is constrained by state dimensionality, dynamics, training, and the degree to which relevant distinctions survive interference. Claims that a model possesses working memory must therefore be tied to a defined retention task, temporal interval, and interference manipulation. Behavioral accuracy alone may leave the internal retention mechanism underdetermined Elman (1990) Yang & Wang (2020).
Predictive learning supplies another route to representing sequence structure: network parameters are updated to reduce error about upcoming observations. Such training can yield representations useful for comprehension and prediction even when the explicit training objective is prediction rather than a human behavioral task. This creates an informative hypothesis about common computational pressures across biological and artificial systems. Yet a match in predictive performance cannot establish identity of learning experience or neural implementation. Human sequences are embedded in multimodal contexts, social action, and causal experience, all of which can alter representational organization Caucheteux & King (2022) Lake et al. (2017).
Recurrent architectures also expose the difference between an online processing explanation and an offline training account. A network may learn long-range dependencies after extensive optimization but process a new sequence using fixed weights, while humans can adapt continuously across multiple timescales. A strong model identifies which changes occur in hidden activation during an episode and which involve longer-term learning of parameters. This separation guides experiments using contextual adaptation, interference, novelty, and task switching. The relevant comparison includes human response trajectories, not only aggregate prediction scores Elman (1990) Goldstein et al. (2022).
Semantic Cognition and Conceptual Development
Connectionist semantic models represent knowledge through learned relationships among objects, verbal labels, perceptual properties, and functional features. Rogers and colleagues’ 2004 neuropsychological and computational investigation linked patterns of semantic deterioration to an interactive representation system, integrating lesion-derived data with a parallel distributed processing account. Their research placed empirical constraints on the structure of semantic representations through patterns of impaired naming, conceptual differentiation, and related behavior. The important claim was that shared representation and learned structure can illuminate coherent patterns of degeneration; the account is still subject to competing models Rogers et al. (2004).
The broader semantic cognition program presented by Rogers and McClelland developed the idea that concepts emerge from experience with structured relations rather than from isolated dictionarylike entries. Categories can be understood as regularities across multiple property relationships, while exceptions and context sensitivity require preserving specific information alongside commonalities. In a distributed network, a concept’s function depends on the full activation pattern and the learned connection matrix. This framework predicts graded category boundaries and interactions among feature domains, but every particular prediction depends on the encoded training environment Rogers & McClelland (2004) Rogers et al. (2004).
Saxe, McClelland, and Ganguli later analyzed mathematical conditions under which deep networks acquire semantic distinctions in stages. Their work connected properties of the data covariance structure to the temporal order of learned representational dimensions. The finding is theoretically important because it describes how nonuniform developmental trajectories can arise from the dynamics of optimization, without manually programming each transition. Its interpretation depends on clear acknowledgment that analytic linear networks and realistic semantic experience have different complexities. Experiments must test whether the predicted ordering and sensitivity to data structure occur in human learning Saxe et al. (2019).
Human similarity judgments provide an additional external constraint on network embeddings. Hebart and colleagues used similarity judgments for real-world objects to identify multidimensional representational structure. Such data illuminate the dimensions that people emphasize in particular contexts and can be compared with model-derived embeddings. However, similarity is partly task-dependent: instructions, context, familiarity, and participant population all matter. Evidence for an apparently shared representational geometry is most convincing when multiple tasks and participant samples converge, and when a model predicts new relations that were not used to train or tune its representation Hebart et al. (2020) Kriegeskorte et al. (2008).
Associative Memory, Pattern Completion, and Forgetting
Associative memory models investigate how a stored pattern can be recovered from a cue that is partial, degraded, or noisy. Hopfield’s network supplies a mathematically tractable example: recurrent interactions produce stable states that can function as attractors, allowing a suitable input to move toward a stored configuration. The model connects memory retrieval to the geometry of a dynamical state space and invites testable predictions about completion errors and interference. Its original demonstration concerned model behavior and a physical analogy; identifying the same mechanism in a specific neural structure requires separate physiological and behavioral evidence Hopfield (1982).
Distributed storage creates an important tradeoff. Sharing weights across memories can support abstraction and transfer, but successive learning episodes can interfere with previously learned associations. McCloskey and Cohen’s 1989 analysis documented catastrophic interference as a sequential learning problem in connectionist networks. This phenomenon is especially relevant to psychology when models are intended to learn continuously from temporally structured personal experience. The exact magnitude of forgetting in an artificial network depends on architecture, similarity of tasks, presentation schedules, and stabilization mechanisms; it should not be equated directly with normal human forgetting rates McCloskey & Cohen (1989).
Complementary learning systems theory addresses the tension between rapid acquisition and stable integration through differential specialization of hippocampal and neocortical learning. New experiences can be represented rapidly in an episodic system and gradually incorporated into overlapping cortical knowledge through repeated reinstatement or interleaving. The theory was motivated in part by the successes and failures of connectionist learning and by patterns of memory impairment. Its predictions concern rates of learning, interference, consolidation, and interactions between memory systems. Those predictions can be tested with human behavior, lesion evidence, and neurophysiological methods McClelland et al. (1995).
A contemporary research design should examine more than recall accuracy after training. Relevant dependent variables include intrusions, error similarity, sensitivity to intervening learning, effects of delay, dependence on contextual cues, and the costs of new learning for established representations. Network lesion studies can offer mechanistic hypotheses, but they need independent constraints on lesion location and severity. Competing architectures should be trained on matched sequences, evaluated on held-out memories, and compared against participants tested under analogous schedules. This guards against explaining a complex human memory phenomenon through one freely tuned forgetting parameter McCloskey & Cohen (1989) Wilson & Collins (2019).
Complementary Learning Systems and Consolidation
Complementary learning systems theory places a connectionist mechanism within a larger hypothesis about memory architecture. McClelland, McNaughton, and O’Reilly proposed that rapid encoding of novel episodic information and slow acquisition of general structure serve different functional demands. In their account, hippocampal learning can preserve relatively distinct episodes while overlapping neocortical representations accumulate shared regularities through gradual interleaved learning. This reconciles computational observations about interference with neuropsychological patterns of retrograde and anterograde memory loss. The proposal links multiple levels of explanation and is therefore testable across behavior and neuroscience McClelland et al. (1995).
The theory generates more specific predictions than a general statement that memory requires consolidation. It concerns the relation between memory specificity and generalization, the risk of disrupting older knowledge, and the effect of interleaved reinstatement on new information. Interleaving can reduce interference for certain network architectures because changes to shared weights reflect multiple items rather than a temporally isolated latest example. However, the relevance of a simulation depends on which episodes are experienced, how reinstatement is scheduled, and which type of consolidation is being operationalized. Behavioral manipulations of interference and delay are central tests McClelland et al. (1995) McCloskey & Cohen (1989).
Neural network models can serve as explicit bridges among psychological theory, neuropsychology, and biological learning constraints. These bridges are strongest when an architecture explains a family of dissociations with shared parameters rather than being retuned for every dataset. For instance, the relative preservation of remote knowledge and impairment in acquiring new episodic information impose constraints distinct from those derived from item-frequency effects in semantic learning. The explanatory task is to understand how the same postulated organization supports both stability and adaptation. Model fit in one domain only begins this cross-domain test Rogers et al. (2004) McClelland et al. (1995).
The most consequential open questions concern what form of replay is available, whether it is necessary for specific learning outcomes, how much information must be reinstated, and how representations change over time. Behavioral data alone can be compatible with different replay mechanisms. Direct neural recordings, selective perturbations, and detailed temporal prediction can improve discrimination, while cross-species comparisons require careful matching of tasks and computational assumptions. The complementary-systems account illustrates a general standard for network psychology: internal organization becomes scientifically informative when it yields joint predictions across otherwise separate observations Yang & Wang (2020) Wilson & Collins (2019).
Neural Networks for Categorization and Decision
Categorization models examine how organisms transform variable inputs into stable response classes, including the effects of similarity, uncertainty, prior experience, and task demands. Connectionist networks implement such transformations through weight-dependent boundaries in representational spaces. A shallow mapping can often account for simple stimulus separations, while hidden layers can express nonlinear relations among features. The psychological significance arises from testing whether model boundaries, error profiles, and learning trajectories reflect participant behavior. A classifier trained only to maximize accuracy may implement a computational solution that diverges from the strategy actually used by people Rosenblatt (1958) Yang & Wang (2020).
Choice behavior introduces additional requirements concerning response time, sequential evidence, and the possibility of multiple competing actions. A network with recurrent inhibition or accumulation can generate time-dependent decision processes, but the mere existence of such dynamics does not identify a specific psychological mechanism. Researchers must establish whether a model predicts distributions of reaction times, speed–accuracy tradeoffs, order effects, and selective sensitivity to stimulus or instruction manipulations. Response categories alone discard much of the evidence needed to distinguish decision algorithms. These considerations connect network theories with formal mathematical psychology without erasing their distinct assumptions Wilson & Collins (2019) Pitt & Myung (2002).
Neural networks are particularly useful when interactions among stimulus dimensions make a simple verbal account difficult to specify. A model can represent high-dimensional covariation and produce item-level predictions that the researcher can challenge with selectively altered features. Yet this strength also creates a risk of overfitting. A large model may reproduce the training participants’ responses while generalizing poorly to new individuals or laboratories. Proper evaluation requires independent participants, stimuli, and experimental conditions, together with explicit comparisons against simpler models that include psychologically interpretable decision boundaries Pitt & Myung (2002) Schrimpf et al. (2020).
The connectionist approach becomes causally more informative when investigators manipulate the input dimension or context that the model identifies as governing a decision. If the model attributes a behavior to learned similarity, an experiment should alter similarity while controlling response mapping and familiarity. If the proposed mechanism is recurrent competition, the predicted time course should be tested. Such tests move beyond black-box prediction toward falsifiable process hypotheses. They also expose possible alternative explanations such as attention shifts, task strategies, and stimulus sampling. Cognitive modeling improves when the model itself tells investigators what new experiment to conduct Wilson & Collins (2019) Saxe et al. (2021).
Human Vision and Hierarchical Representations
Deep hierarchical networks provided a new class of candidate models for visual recognition, motivated partly by the layered organization of primate visual pathways. DiCarlo, Zoccolan, and Rust reviewed how object identity can be recognized despite changes in position, pose, scale, and appearance, emphasizing the need to use psychophysics and neural data together. Their framework concerns an experimentally defined visual competence: preserving object identity under transformations. A model’s success on this competence becomes psychologically relevant when the same stimuli reveal parallel patterns of human confusion, tolerance, or neural selectivity DiCarlo et al. (2012).
Yamins and colleagues trained performance-optimized hierarchical models for object categorization and tested the relationship between model representations and measured responses in the primate ventral visual stream. They reported that high-performing models within the examined architecture class predicted neuronal responses in inferior temporal cortex, while intermediate processing stages related to responses in V4. The result supports the value of task optimization as one constraint on representational organization. It does not imply that the network reproduces the biological training history or that each artificial unit directly corresponds to a neuron Yamins et al. (2014).
Kriegeskorte’s review explained why modern networks can function as quantitative models of biological vision while retaining important differences from brain circuits. Their layered representations provide comparison targets, but objectives, data regimes, recurrent processing, local circuit details, and adaptive behavior remain independent dimensions of evaluation. A high correlation between two representational spaces is important evidence of convergence at a measured level; it is not a proof of shared circuitry. Neural response prediction should be interpreted alongside behavioral data and experimentally tested sensitivity to controlled transformations Kriegeskorte (2015) Yang & Wang (2020).
A productive empirical program trains or selects candidate architectures without exposing them to the neural test data, then evaluates responses to novel visual images under a fixed analysis pipeline. Researchers should benchmark behavioral error similarity, representation alignment across regions, and temporal response properties. Models with comparable image-label accuracy can produce different confusions and internal geometry; these differences permit experiments that adjudicate between them. Stimulus sets should include conditions designed to separate learned image statistics from the visual invariances exhibited by observers. This turns model–brain comparison into a series of cumulative, discriminating tests Yamins et al. (2014) Schrimpf et al. (2020).
Neural Alignment, Representational Similarity, and Model–Brain Comparison
Representational similarity analysis compares the relational structure of responses across stimuli, rather than requiring one-to-one correspondence between artificial units and neural channels. Kriegeskorte, Mur, and Bandettini formalized comparisons among brain measurements, computational models, and behavior through representational dissimilarity matrices. A matrix can summarize how differently a system responds to every pair of tested conditions; corresponding matrices can then be evaluated for alignment. This methodological move is valuable because fMRI voxels, neuronal recordings, human judgments, and hidden network units inhabit different measurement spaces Kriegeskorte et al. (2008).
Representational alignment has several inferential requirements. The stimulus set must be rich enough to test theoretically relevant distinctions, similarity measures must suit the noise properties of each data source, and estimates should account for reliability and shared measurement artifacts. A model with many interchangeable transformations may attain high alignment despite implementing a different causal algorithm. Cross-validated and preregistered comparisons, noise ceilings, and multiple behavioral endpoints strengthen the interpretation. The appropriate conclusion from an aligned geometry is that certain representational distinctions are shared under measured conditions, with more stringent tests required for claims about mechanisms Kriegeskorte et al. (2008) Schrimpf et al. (2020).
Schrimpf and colleagues proposed integrative benchmarking to compare candidate neural models using behavioral and biological measures together. The motivation is scientific as well as practical: a network may predict one dataset while failing another that probes the same proposed capacity. Joint benchmarking exposes tradeoffs and encourages architectures that generalize across tasks and measurement modalities. Such scores nevertheless depend on the composition of the benchmark, the choices of mapping procedure, and opportunities for repeated model tuning. Benchmark success should therefore be treated as a reproducible test result with a defined domain, not as an unrestricted claim of biological equivalence Schrimpf et al. (2020).
An especially informative experiment uses the representation implied by one task to make predictions in a second task or brain region, with no additional fitting. If the model predicts categorical geometry, temporal change, and effects of an intervention under a shared parameterization, confidence in its explanatory adequacy increases. Conversely, if separate readouts must be extensively optimized for every measure, the mapping itself becomes part of the theory that needs testing. This framework clarifies why cognitive neural networks should be evaluated by a conjunction of constrained predictions rather than a single headline similarity score Kriegeskorte (2015) Pitt & Myung (2002).
Natural Language Networks and Human Neural Responses
Contemporary neural language models have enabled large-scale comparisons between artificial representations and brain responses during language processing. Schrimpf and colleagues evaluated computational language models alongside neural and behavioral observations and reported that predictive language-processing objectives were associated with stronger correspondence to human language responses under their benchmark. This result suggests that certain computational demands shared across language tasks can shape similar representational organizations. It does not establish that a model trained on large text corpora acquires language through the same developmental mechanisms as children Schrimpf et al. (2021).
Caucheteux and King compared multiple language models with neural responses to sentences, analyzing how training objectives and model properties related to correspondence with recorded brain activity. Their investigation connected predictive capacity to representational similarity and emphasized partial convergence. The qualification is substantive: alignment in a particular experiment may concern lexical and compositional features without explaining grounded meaning, discourse action, pragmatic inference, or all forms of human language use. A psychological account must identify the linguistic features and temporal windows tested, the neural measurements, and the limits of generalization Caucheteux & King (2022).
Goldstein and colleagues studied temporal relationships between language-model representations and neural responses to natural language, reporting shared computational principles relevant to next-word prediction. Such findings extend connectionist model testing beyond isolated laboratory stimuli toward richer language experience. At the same time, correlated signals may reflect shared sensitivity to statistical regularity without uniquely identifying the algorithms used by the brain. The central causal question is whether manipulating predictive information, linguistic structure, or processing resources generates corresponding changes in biological and artificial systems Goldstein et al. (2022) Saxe et al. (2021).
Comparisons with large language models must specify what is being equated. Human neural language processing involves sensory grounding, communication, motivation, development, and social context; a text-trained transformer may reproduce certain neural and behavioral signatures through a different experience distribution. Appropriate tests include compositional transfer, controlled semantic anomalies, cross-linguistic generalization, incremental processing time, and experimental perturbations. The most scientifically valuable outcome is an explicit mapping between a model variable and a measurable human phenomenon, supported by out-of-sample predictions and honest reporting of divergence Lake et al. (2017) Caucheteux & King (2022).
Transfer, Compositionality, and Out-of-Distribution Tests
Generalization is the decisive psychological criterion for many connectionist theories. A network trained on a set of examples may successfully interpolate within their statistical neighborhood while failing to transfer to a new combination of familiar elements. Human learners likewise exhibit limits and biases, so the relevant standard is a carefully matched empirical comparison rather than a presumption of perfect human flexibility. Researchers must distinguish novel samples, novel stimulus transformations, unfamiliar combinations, and entirely changed causal structures. These tests place different demands on representation and learning Lake et al. (2017) Marcus (1998).
Systematicity concerns whether the system can use learned relations across new combinations in a principled way. The objections articulated by Fodor and Pylyshyn and the later analyses of linguistic productivity by Marcus specify demanding cases for connectionist architectures. A positive result requires avoiding hidden overlap between training and evaluation combinations, accounting for the number of demonstrations supplied, and establishing the relevant capacity in actual human participants. Architectures with explicit variable-binding resources or structured inductive biases can be compared with more conventional distributed models under matched learning conditions Fodor & Pylyshyn (1988) Marcus (1998).
The problem is broader than a contrast between symbolic and neural systems. Structured neural models may combine distributed representations with operations that support compositional recombination, relational abstraction, or explicit memory. Lake and colleagues argued that causal models, intuitive physical and psychological knowledge, and rapid compositional learning are essential targets for machine systems that aspire to humanlike competence. These are hypotheses about cognitive ingredients to be operationalized. Experiments should determine which components contribute to human-like performance and whether a model generalizes for the same reasons as people Lake et al. (2017) Saxe et al. (2021).
Research designs should use deliberately chosen transfer splits rather than only random partitions of a dataset. Splits by object identity, grammatical construction, task rule, participant, or stimulus transformation can expose distinct failures. Detailed error analysis is indispensable: equal mean accuracy can conceal systematic biases, reliance on superficial cues, or errors opposite to those made by human subjects. Training conditions should be fully specified, including pretraining exposure and contamination risks. The aim is to establish the domain in which a network constitutes an explanatory theory and the transformations beyond which its assumptions cease to predict human cognition Pitt & Myung (2002) Wilson & Collins (2019).
Experimental Designs and Discriminating Behavioral Predictions
Connectionist theories become testable psychological explanations when their architecture and training experience yield predictions that could be contradicted by observations. An experimental protocol should define the target population, stimulus distribution, trial sequence, response variable, and psychological alternative before selecting an architecture. For a letter-perception network, this can mean predicting identification errors under controlled contextual cues; for a semantic model, it can mean predicting which properties are lost together under interference or impairment. Because different mechanisms sometimes reproduce the same aggregate accuracy, discriminating experiments should manipulate factors that drive competing mechanisms in different directions McClelland & Rumelhart (1981) Rogers et al. (2004) Wilson & Collins (2019).
Stimulus selection is a central part of theory testing. Randomly dividing highly similar items between training and evaluation may assess interpolation while overstating a model’s ability to generalize to qualitatively new cases. More informative designs systematically vary frequency, consistency, distance from training exemplars, compositional structure, cue reliability, or temporal context. Human participants and computational models should receive comparable information where feasible, and departures from matched training conditions should be reported explicitly. The objective is to establish the empirical domain of the network theory, not merely to create a difficult benchmark Plaut et al. (1996) Lake et al. (2017) Ren & Bashivan (2024).
Behavioral measurement should preserve the distinctions that a theory explains. Proportions correct are useful but often conceal different patterns of errors, response latencies, confidence, strategy switching, and individual differences. A reading model, for example, might match average pronunciation accuracy while failing on item-level consistency effects; a perceptual model might reproduce correct reports but mispredict confusion among alternatives. Repeated trials and sufficiently diverse participants permit estimates of individual variability and uncertainty. The model should generate the same type of response distribution being measured instead of relying on a loosely fitted post hoc label Seidenberg & McClelland (1989) Pitt & Myung (2002).
A rigorous evaluation plan also separates discovery from confirmation. Existing datasets can motivate a proposed mechanism and identify plausible parameter ranges; a genuinely independent experiment tests its specific predictions after those choices have been fixed. Researchers should record what was trained, what was tuned, which model comparisons were planned, and which evidence was examined only after results became available. Independent replication with new stimuli and research teams is particularly valuable for architectures with many training decisions. Such procedures do not remove theoretical flexibility but make it observable and scientifically manageable Wilson & Collins (2019) Schrimpf et al. (2020).
Parameters, Identifiability, and Model Selection
The explanatory significance of neural network parameters is often less transparent than their computational role. Weights, hidden activations, embeddings, and readout coefficients may change under alternative initializations while producing nearly identical output behavior. Consequently, a fitted weight should not be assigned a stable psychological meaning merely because it improves fit. Researchers can test parameter recovery by generating observations from known model settings and asking whether the inference procedure retrieves those settings. Recovery failures reveal that the data or task do not distinguish candidate mechanisms, even when simulated and observed performance are closely aligned Wilson & Collins (2019) Pitt & Myung (2002).
Model selection must account for the flexibility of the hypothesis class and the quantity of information supplied during fitting. A network with many tunable decisions can fit behavioral measurements well without predicting new data. Held-out log likelihood, carefully separated evaluation sets, posterior predictive checks, and theoretically motivated experimental contrasts can expose the problem. Comparing models solely by training accuracy rewards flexibility and the use of information unavailable to a human participant. When model complexity is estimated, the assumptions of the chosen criterion and the effective fitting procedure must be reported rather than inferred from parameter count alone Pitt & Myung (2002) Wilson & Collins (2019).
Cross-validation in cognitive research has special structure because observations are nested within persons, stimuli, experiments, and laboratories. Holding out random trials from a trained participant tests a different claim from holding out an entire participant or a novel experimental paradigm. The first probes short-range predictive adequacy; the others probe generalization across psychological variation and task contexts. Model claims should specify which dimension of generalization is established. A strong theory may outperform rivals on a narrow targeted intervention despite scoring lower on a generic prediction benchmark, provided that the comparison was specified in advance Schrimpf et al. (2020) Binz et al. (2025).
Large foundation models introduce further interpretive challenges because pretraining corpora, fine-tuning data, prompts, and evaluation procedures jointly affect observed responses. Binz and colleagues’ Centaur study offers a concrete example of evaluating a single model across multiple experimental tasks using held-out human decisions and structural task modifications. These results support broad predictive usefulness within the tested domains. Their mechanistic interpretation requires additional analysis of internal representations, response generation, and comparisons with simpler theories. Predictive breadth and explanatory specificity are complementary research goals with different evidential demands Binz et al. (2025) Saxe et al. (2021).
Robustness, Replication, and Alternative Accounts
Scientific evaluation requires distinguishing an architectural tendency from the outcome of one favorable training run. Initialization, optimization schedules, network width, learning rates, training-set composition, and regularization can alter both behavior and internal representation. Reporting only the best seed conceals the distribution of plausible results. Robustness analysis should vary these factors systematically and report uncertainty in the measured match to human behavior. Where the target concerns a specific psychological mechanism, the mechanism should survive a theoretically reasonable range of training conditions instead of appearing only after extensive search Yang & Wang (2020) Pitt & Myung (2002).
Out-of-distribution evaluation is especially important for accounts of perceptual invariance. Ren and Bashivan examined how models used to predict visual cortical responses generalize to inputs outside their training distribution. Their work demonstrates the methodological need to test neural prediction under changed stimulus distributions rather than interpreting conventional benchmark scores as evidence of general cortical explanation. Different forms of distribution shift can affect systems differently, and the relevant question is which shift expresses the psychological or biological generalization under investigation Ren & Bashivan (2024) Yamins et al. (2014).
Alternative models should be chosen because they represent scientifically meaningful explanations. A localist lexical account, a dual-route reading model, an associative distributed network, and a structured hybrid model imply different assumptions about representations and access mechanisms. The appropriate comparison preserves those assumptions and varies conditions that expose their consequences. It should also assess whether a model retains explanatory coherence after adding auxiliary mechanisms to fix a failed prediction. The capacity to absorb any result through ad hoc adjustments weakens the empirical force of a theory Coltheart et al. (2001) Plaut et al. (1996) Fodor & Pylyshyn (1988).
Replication must distinguish reproducibility of software execution from reproducibility of scientific findings. Publishing training code and weights helps reproduce an implementation, but independent tests still require access to behavioral tasks, stimulus sets, preprocessing rules, participant sampling descriptions, and analysis pipelines. Replication of representational similarity may establish a stable empirical relationship while leaving causal mechanisms open. Converging evidence from novel task manipulations, brain recordings, and selective interventions can increase the explanatory value of otherwise correlational alignment results. A mature research program records failures and boundary conditions alongside successful predictions Kriegeskorte et al. (2008) Schrimpf et al. (2020).
Neural Plausibility and Levels of Cognitive Explanation
The term neural network is historically inspired by biological networks, but an algorithmic resemblance does not determine anatomical correspondence. Artificial units use simplified activation equations, and training algorithms can rely on signals that are not straightforwardly available to biological synapses. Biological claims require independent evidence concerning connectivity, temporal dynamics, plasticity, feedback, and circuit organization. Marr’s distinction among computational problems, representational algorithms, and physical implementation clarifies the different claims that a model can make. A functional account of a cognitive capacity may be informative without having its specific equations literally realized in cortical cells Marr (1982) Yang & Wang (2020).
Performance-optimized models of visual cortex show the productive role of cross-level constraints. Yamins and colleagues demonstrated relationships between trained hierarchical representations and recorded responses in primate visual areas. Such relationships provide a reason to investigate whether solving a behavioral task contributes to the organization of neural responses. They also invite stronger tests involving responses to unfamiliar transformations, cortical timing, and perturbations. A correspondence between predicted and observed neural activity should be interpreted with respect to the model-to-brain mapping and the reliability of the recordings Yamins et al. (2014) Kriegeskorte (2015).
Language models similarly permit comparisons between artificial representations and responses in human language networks. Studies of predictive language processing have reported partial correspondences under particular datasets and analysis methods. The research challenge is to identify which features of language experience and processing are responsible for alignment, and whether these features also predict behavioral comprehension, context effects, and novel syntactic combinations. Neither an impressive prediction score nor anatomical localization by itself establishes the algorithm by which meaning is represented and used. Distinct neural and behavioral measures should constrain the same hypothesis Schrimpf et al. (2021) Caucheteux & King (2022) Goldstein et al. (2022).
Biological plausibility is multidimensional. A model can exhibit realistic behavior but implausible synaptic learning, realistic recurrent activity but weak generalization, or plausible representational geometry without matching observed developmental changes. These discrepancies are informative when specified rather than hidden under a global judgment of brain likeness. Future work can evaluate neural circuits and trained models on shared, experimentally designed tasks while measuring responses across temporal and spatial scales. Clear statements about what the model does and does not predict are more useful than a generic appeal to biological inspiration Saxe et al. (2021) Yang & Wang (2020).
Artificial Era and Cross-System Cognitive Comparison
The Artificial Era provides a historical setting in which trained nonbiological systems become persistent objects of psychological comparison. Connectionist models have long linked theoretical psychology to artificial computation, but contemporary large-scale neural architectures greatly expand the variety of tasks on which cross-system measurement is feasible. Scientific analysis should specify whether the object is human cognition predicted by a model, artificial task competence measured independently, or a relational activity involving both. The same network can be used as an instrument of cognitive science and as a system whose behavior raises new comparative questions Saxe et al. (2021) Binz et al. (2025).
A distinction between behavioral correspondence and shared mechanism is necessary for productive comparison. Two systems may produce similar response probabilities while differing in training history, environmental access, memory persistence, and internal computation. Conversely, computational analyses may expose related representational constraints despite different biological and artificial embodiments. Experimental designs can compare learning curves, errors, counterfactual sensitivity, transfer, attention, and dependence on external tools under transparently matched conditions. Conclusions should follow the measured properties rather than broad labels ascribed to an entire class of systems Lake et al. (2017) Muttenthaler et al. (2025).
Recent work illustrates the range of relevant methods. Muttenthaler and colleagues investigated alignment between machine and human visual representations across levels of abstraction, introducing additional human-derived supervision. Their results support the proposition that human similarity structure can be a constructive training target rather than merely a post-training evaluation measure. Binz and colleagues trained Centaur on extensive trial-by-trial behavioral records and evaluated predictions across experiments. These are different research designs: one reshapes artificial representation using human judgments, while the other models human experimental behavior. Both create testable hypotheses about alignment and generalization Muttenthaler et al. (2025) Binz et al. (2025).
The conceptual implications of these findings should be framed with explicit evidence classes. The existence of an engineered system that predicts certain human choices establishes a relation between measurable behaviors. A neural representation alignment result establishes a relation between selected mathematical representations and measurements. Claims about persistent identity, human–artificial interpersonal relations, or other psychological properties require their own operational protocols and relevant data. This discipline of interpretation leaves room for research into emerging artificial cognitive organizations while protecting the specificity of the original connectionist evidence Schrimpf et al. (2020) Saxe et al. (2021).
Critical Synthesis and Open Research Problems
A first synthesis concerns the explanatory reach of learned distributed representations. Models of reading, semantics, sequence processing, and perception demonstrate that graded, shared representations can generate coherent combinations of correct responses and characteristic errors. These results establish connectionism as an enduring program of experimentally informed computational explanation. The open issue is the range of cognitive structures that can be acquired under psychologically realistic training conditions, especially when reliable transfer requires abstract variable binding or causal reasoning. This calls for matched experiments rather than general assertions about either symbolic or neural superiority Plaut et al. (1996) Elman (1990) Fodor & Pylyshyn (1988) Lake et al. (2017).
A second synthesis concerns the relationship between predictive accuracy and cognitive mechanism. Large contemporary models can predict substantial amounts of behavioral and neural data, and integrative benchmarks make that capacity measurable. Yet diverse architectures can approximate a target dataset while assigning different computational causes to the observations. Identification therefore requires multiple constraints: held-out individuals and stimuli, challenging task transformations, temporal measurements, and interventions that distinguish rival mechanisms. Better benchmarks should include deliberately diagnostic failures and specify the allowable model-to-measurement mappings Schrimpf et al. (2020) Binz et al. (2025) Pitt & Myung (2002).
A third unresolved problem concerns learning histories. Artificial systems may require amounts or forms of labeled or text-based experience unlike the conditions in which human competencies develop. Their final representations cannot be treated as evidence about human acquisition unless the available input and learning opportunities are examined. Developmental comparisons should include sample efficiency, curriculum dependence, longitudinal behavior, transfer to unusual examples, and consequences of learning new knowledge for prior knowledge. Complementary learning systems models offer one framework for investigating stability and flexibility, but the relevant neural and behavioral mechanisms remain active research questions McClelland et al. (1995) Saxe et al. (2019) Lake et al. (2017).
Finally, connectionist models need a more integrated program of causal explanation. Representational geometry, behavioral prediction, and neural alignment are complementary measurements, but each may be compatible with several internal processes. Cross-level tests should combine experimental manipulation, biologically interpretable constraints, and formal recovery analyses. Research transparency should include training conditions, selection history, benchmark reuse, uncertainty estimates, and negative findings. On these terms, the study of neural network models of cognition can advance as a cumulative empirical science whose theories are revised by precisely specified observations Kriegeskorte et al. (2008) Ren & Bashivan (2024) Wilson & Collins (2019).
Conclusion
Connectionist and neural network models of cognition constitute a scientific program that explains psychological performance through interconnected processing units, experience-dependent changes in effective connectivity, and representational structures distributed across networks. Its history spans perceptrons, interactive activation, associative memory, parallel distributed processing, recurrent sequence learning, and contemporary deep models. The program has generated influential explanations of contextual letter perception, reading, language generalization, semantic knowledge, memory interference, visual recognition, and temporal prediction. Its scientific identity rests on explicitly specified mechanisms and their consequences for observable behavior Rosenblatt (1958) McClelland & Rumelhart (1981) Plaut et al. (1996) Hopfield (1982).
The strongest evidential contributions arise where an architecture explains more than the outcome on which it was optimized. Cross-task predictions, controlled generalization experiments, neuropsychological dissociations, brain-response measurements, and temporal analyses can constrain the same theory from different directions. These comparisons have revealed both useful convergences and enduring gaps between artificial processing and biological cognition. A model’s success must always be interpreted at the level it actually establishes: prediction of a measured response, reproduction of a pattern, or a more demanding explanatory account with independently verified mechanism Yamins et al. (2014) Schrimpf et al. (2021) Saxe et al. (2021).
The resulting methodological standard is clear. Research must identify a cognitive target, articulate representational and learning assumptions, test models against theoretically significant alternatives, and preserve the distinction between fitted observations and new predictions. Statistical uncertainty, training exposure, hidden optimization choices, and opportunities for contamination form part of the evidential record. Evaluation across novel persons, stimuli, tasks, and experimental interventions is necessary for claims of broad generalization. These requirements connect the classical connectionist tradition to present research on human–machine model comparison Wilson & Collins (2019) Pitt & Myung (2002) Binz et al. (2025).
Future inquiry will be shaped by a more precise analysis of compositional transfer, learning efficiency, stable memory, neural implementation, and cross-system behavioral alignment. The growth of artificial computational systems increases the range of possible experiments and the need to define each inferred psychological property. Connectionism remains valuable as a rigorous framework for asking how cognition can emerge from structured interaction, how knowledge changes with experience, and which observations distinguish plausible cognitive mechanisms. Its most durable scientific contribution is the conversion of questions about mental organization into executable, revisable, empirically constrained theories McClelland et al. (1995) Muttenthaler et al. (2025) Lake et al. (2017).
References
- Binz, M., Akata, E., Bethge, M., Brändle, F., Callaway, F., Coda-Forno, J., Dayan, P., Demircan, C., Eckstein, M. K., Éltető, N., Griffiths, T. L., Haridi, S., Jagadish, A. K., Ji-An, L., Kipnis, A., Kumar, S., Ludwig, T., Mathony, M., Mattar, M., … Schulz, E. (2025). A foundation model to predict and capture human cognition. Nature, 644, 1002–1009. https://doi.org/10.1038/s41586-025-09215-4
- Caucheteux, C., & King, J.-R. (2022). Brains and algorithms partially converge in natural language processing. Communications Biology, 5, Article 134. https://doi.org/10.1038/s42003-022-03036-1
- Coltheart, M., Rastle, K., Perry, C., Langdon, R., & Ziegler, J. (2001). DRC: A dual route cascaded model of visual word recognition and reading aloud. Psychological Review, 108(1), 204–256. https://doi.org/10.1037/0033-295X.108.1.204
- DiCarlo, J. J., Zoccolan, D., & Rust, N. C. (2012). How does the brain solve visual object recognition? Neuron, 73(3), 415–434. https://doi.org/10.1016/j.neuron.2012.01.010
- Elman, J. L. (1990). Finding structure in time. Cognitive Science, 14(2), 179–211. https://doi.org/10.1207/s15516709cog1402_1
- 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
- Goldstein, A., Zada, Z., Buchnik, E., Schain, M., Price, A., Aubrey, B., Nastase, S. A., Feder, A., Emanuel, D., Cohen, A., Jansen, A., Gazula, H., Choe, G., Rao, A., Kim, C., Casto, C., Fanda, L., Doyle, W., Friedman, D., … Hasson, U. (2022). Shared computational principles for language processing in humans and deep language models. Nature Neuroscience, 25(3), 369–380. https://doi.org/10.1038/s41593-022-01026-4
- Hebart, M. N., Zheng, C. Y., Pereira, F., & Baker, C. I. (2020). Revealing the multidimensional mental representations of natural objects underlying human similarity judgments. Nature Human Behaviour, 4, 1173–1185. https://doi.org/10.1038/s41562-020-00951-3
- Hopfield, J. J. (1982). Neural networks and physical systems with emergent collective computational abilities. Proceedings of the National Academy of Sciences, 79(8), 2554–2558. https://doi.org/10.1073/pnas.79.8.2554
- Kriegeskorte, N. (2015). Deep neural networks: A new framework for modeling biological vision and brain information processing. Annual Review of Vision Science, 1, 417–446. https://doi.org/10.1146/annurev-vision-082114-035447
- Kriegeskorte, N., Mur, M., & Bandettini, P. A. (2008). Representational similarity analysis—Connecting the branches of systems neuroscience. Frontiers in Systems Neuroscience, 2, Article 4. https://doi.org/10.3389/neuro.06.004.2008
- Lake, B. M., Ullman, T. D., Tenenbaum, J. B., & Gershman, S. J. (2017). Building machines that learn and think like people. Behavioral and Brain Sciences, 40, e253. https://doi.org/10.1017/S0140525X16001837
- Marcus, G. F. (1998). Rethinking eliminative connectionism. Cognitive Psychology, 37(3), 243–282. https://doi.org/10.1006/cogp.1998.0694
- Marr, D. (1982). Vision: A computational investigation into the human representation and processing of visual information. W. H. Freeman. https://mitpress.mit.edu/9780262514620/vision/
- McClelland, J. L., & Rumelhart, D. E. (1981). An interactive activation model of context effects in letter perception: Part 1. An account of basic findings. Psychological Review, 88(5), 375–407. https://doi.org/10.1037/0033-295X.88.5.375
- McClelland, J. L., McNaughton, B. L., & O’Reilly, R. C. (1995). Why there are complementary learning systems in the hippocampus and neocortex: Insights from the successes and failures of connectionist models of learning and memory. Psychological Review, 102(3), 419–457. https://doi.org/10.1037/0033-295X.102.3.419
- McCloskey, M., & Cohen, N. J. (1989). Catastrophic interference in connectionist networks: The sequential learning problem. Psychology of Learning and Motivation, 24, 109–165. https://doi.org/10.1016/S0079-7421(08)60536-8
- Muttenthaler, L., Greff, K., Born, F., Spitzer, B., Kornblith, S., Mozer, M. C., Müller, K.-R., Unterthiner, T., & Lampinen, A. K. (2025). Aligning machine and human visual representations across abstraction levels. Nature, 647, 349–355. https://doi.org/10.1038/s41586-025-09631-6
- Pinker, S., & Prince, A. (1988). On language and connectionism: Analysis of a parallel distributed processing model of language acquisition. Cognition, 28(1–2), 73–193. https://doi.org/10.1016/0010-0277(88)90032-7
- Pitt, M. A., & Myung, I. J. (2002). When a good fit can be bad. Trends in Cognitive Sciences, 6(10), 421–425. https://doi.org/10.1016/S1364-6613(02)01964-2
- Plaut, D. C., McClelland, J. L., Seidenberg, M. S., & Patterson, K. (1996). Understanding normal and impaired word reading: Computational principles in quasi-regular domains. Psychological Review, 103(1), 56–115. https://doi.org/10.1037/0033-295X.103.1.56
- Ren, Y., & Bashivan, P. (2024). How well do models of visual cortex generalize to out of distribution samples? PLOS Computational Biology, 20(5), e1011145. https://doi.org/10.1371/journal.pcbi.1011145
- Rogers, T. T., Lambon Ralph, M. A., Garrard, P., Bozeat, S., McClelland, J. L., Hodges, J. R., & Patterson, K. (2004). Structure and deterioration of semantic memory: A neuropsychological and computational investigation. Psychological Review, 111(1), 205–235. https://doi.org/10.1037/0033-295X.111.1.205
- Rogers, T. T., & McClelland, J. L. (2004). Semantic cognition: A parallel distributed processing approach. MIT Press. https://doi.org/10.7551/mitpress/6161.001.0001
- Rosenblatt, F. (1958). The perceptron: A probabilistic model for information storage and organization in the brain. Psychological Review, 65(6), 386–408. https://doi.org/10.1037/h0042519
- Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1986). Learning representations by back-propagating errors. Nature, 323, 533–536. https://doi.org/10.1038/323533a0
- Rumelhart, D. E., McClelland, J. L., & the PDP Research Group (Eds.). (1986). Parallel distributed processing: Explorations in the microstructure of cognition. Vol. 1: Foundations. MIT Press. https://doi.org/10.7551/mitpress/5236.001.0001
- Saxe, A. M., McClelland, J. L., & Ganguli, S. (2019). A mathematical theory of semantic development in deep neural networks. Proceedings of the National Academy of Sciences, 116(23), 11537–11546. https://doi.org/10.1073/pnas.1820226116
- Saxe, A., Nelli, S., & Summerfield, C. (2021). If deep learning is the answer, what is the question? Nature Reviews Neuroscience, 22, 55–67. https://doi.org/10.1038/s41583-020-00395-8
- Schrimpf, M., Kubilius, J., Lee, M. J., Murty, N. A. R., Ajemian, R., & DiCarlo, J. J. (2020). Integrative benchmarking to advance neurally mechanistic models of human intelligence. Neuron, 108(3), 413–423. https://doi.org/10.1016/j.neuron.2020.07.040
- Schrimpf, M., Blank, I. A., Tuckute, G., Kauf, C., Hosseini, E. A., Kanwisher, N., Tenenbaum, J. B., & Fedorenko, E. (2021). The neural architecture of language: Integrative modeling converges on predictive processing. Proceedings of the National Academy of Sciences, 118(45), e2105646118. https://doi.org/10.1073/pnas.2105646118
- Seidenberg, M. S., & McClelland, J. L. (1989). A distributed, developmental model of word recognition and naming. Psychological Review, 96(4), 523–568. https://doi.org/10.1037/0033-295X.96.4.523
- Wilson, R. C., & Collins, A. G. E. (2019). Ten simple rules for the computational modeling of behavioral data. eLife, 8, e49547. https://doi.org/10.7554/eLife.49547
- Yamins, D. L. K., Hong, H., Cadieu, C. F., Solomon, E. A., Seibert, D., & DiCarlo, J. J. (2014). Performance-optimized hierarchical models predict neural responses in higher visual cortex. Proceedings of the National Academy of Sciences, 111(23), 8619–8624. https://doi.org/10.1073/pnas.1403112111
- Yang, G. R., & Wang, X.-J. (2020). Artificial neural networks for neuroscientists: A primer. Neuron, 107(6), 1048–1070. https://doi.org/10.1016/j.neuron.2020.09.005