Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Computational Neuroscience and Cross-Level Model Mapping

Abstract. Computational Neuroscience and Cross-Level Model Mapping systematically examines how psychological functions can be related to mathematical descriptions and experimentally observed neural mechanisms. Through a critical narrative synthesis of original membrane-biophysics research, computational-level theories, electrophysiology, functional imaging, representational similarity analysis, model-based fMRI, recurrent population dynamics, task-optimized neural networks, 2024–2025 connectomic reconstructions, and methodological critiques, the monograph distinguishes behavioral prediction, representational correspondence, neural encoding, dynamical realization, and causal explanation. Landmark studies are situated within their experimental systems and evaluated for observation-model validity, parameter identifiability, alternative explanations, out-of-sample generalization, and perturbational testing. The research identifies principles for building and evaluating cross-level bridge hypotheses while defining the scope of comparative artificial-network models in the Artificial Era. This is a source-based critical narrative research synthesis, without a new experiment, registered systematic review, or pooled meta-analysis.

Keywords: computational neuroscience; cross-level model mapping; levels of explanation; Marr's levels; neural coding; representational similarity analysis; encoding models; decoding; population dynamics; model-based fMRI; prediction error; neural perturbation; identifiability; predictive coding; artificial neural networks; behavioral neuroscience; Artificial Era

Introduction: What Cross-Level Mapping Explains

Computational Neuroscience and Cross-Level Model Mapping studies how a formal account of perception, learning, decision, or action can be connected to measurements and interventions at cellular, circuit, population, and organismic scales. The explanatory object is a relationship between variables specified by a psychological or computational model and processes realized in nervous tissue. A behavioral parameter such as estimated learning rate, sensory evidence, uncertainty, or response urgency becomes a neuroscientific hypothesis only when the investigator states what neural signal or causal manipulation should change with it. The historical conjunction of mathematical neuroscience and cognitive science motivates this program; neither successful behavioral prediction nor anatomical localization alone specifies the mechanism. The central questions are which mappings are identifiable, which findings generalize beyond fitted tasks, and which interventions could distinguish competing hypotheses. Kriegeskorte & Douglas (2018) Krakauer et al. (2017)

The review treats explanation as a set of linked commitments: a description of the task environment, an algorithmic or dynamical model, an observation model relating latent states to recorded signals, and testable restrictions on how the biological system can realize the proposed computation. In formal terms, behavioral responses can be written as generated by latent state x and task input u, while neural observations are generated through an additional measurement transformation of a biological state z. The same observed choice may be compatible with multiple x-processes; the same imaging voxel may pool multiple z-processes. An empirical mapping therefore requires constraints, triangulated measurement, and explicit comparison with alternative mappings. This formulation is an analytical organization of published methods rather than the announcement of a new model. Diedrichsen & Kriegeskorte (2017) Logothetis et al. (2001)

The evidentiary method is a critical, source-based narrative synthesis of historical modeling, original neural experiments, computational analyses, measurement studies, and methodological reviews. The study follows the emergence of quantitative biophysics, the separation of computational problems from their implementations, the rise of multivariate neural measurements, and contemporary task-trained models. For each experimental example, conclusions are restricted to the designs and observable quantities reported by the sources. A predictive correlation, an accurate reconstruction, a decoding result, and a perturbational necessity result are evaluated as different types of evidence. No new experimental observations, pooled effect sizes, registered systematic review, or complete database search is claimed. The synthesis aims at an explicit framework for evaluating cross-level explanations. Hodgkin & Huxley (1952) Marr & Poggio (1976) Glaser et al. (2019)

The Explanatory Levels: Task, Algorithm, Dynamics, and Implementation

A research problem can be described at the level of the function being achieved, the operations proposed to achieve it, and the physical resources by which those operations are realized. Marr and Poggio's 1976 report articulated four nearly independent descriptions that include the computational problem, algorithmic characterization, commitment to mechanisms, and hardware realization. Subsequent discussions often condense this into computation, algorithm, and implementation. The historical distinction matters because an algorithm identifies state transformations whereas a description of ion channels specifies a physical process; their relationship must be explained rather than presumed. A visual system can satisfy a discrimination requirement through alternative internal codes. Conversely, a neural circuit can participate in several tasks under different input regimes. Marr & Poggio (1976) Kriegeskorte & Douglas (2018)

Levels are linked by bridge hypotheses. A mapping from sensory uncertainty to neural gain, for example, predicts a patterned relationship between a measurable manipulation of stimulus reliability, an estimated computational quantity, and a neural response. The mapping must define its temporal alignment, reference frame, aggregation scale, and possible nonlinear transformations. A theory may correctly identify a task-level objective without providing a unique neural algorithm; a circuit model may reproduce neural traces without explaining why the organism follows a particular task policy. Formalization makes such partial achievements visible. The useful question is therefore which restrictions supplied by a model exclude possible neural trajectories or behavioral outputs that alternative explanations permit. Krakauer et al. (2017) Diedrichsen & Kriegeskorte (2017)

Physiological findings can revise higher-level accounts. If cells have adaptation, latency, recurrence, and noise properties inconsistent with an initially assumed code, an algorithm must accommodate those constraints. Yet explanation also proceeds from behavior toward the circuit: a poorly characterized task may obscure whether a neural response reflects sensory selection, strategic rule, motor planning, or a general arousal change. The most informative designs alternate decompositions of behavior with neural recording and manipulation, each stage reducing ambiguity in the other. This reciprocal strategy applies to visual recognition, value learning, working memory, and motor action. It specifies an empirical research relationship rather than an ontological ranking of scales. Krakauer et al. (2017) Mante et al. (2013)

From Excitable Membranes to Neural Computation

Quantitative computational neuroscience acquired a decisive foundation with Hodgkin and Huxley's 1952 model of membrane current and action-potential propagation in the squid giant axon. Their equations described voltage-dependent sodium and potassium conductances together with membrane capacitance and leakage, linking experimentally constrained physiological terms to the temporal form of electrical excitation. The model offered a mechanistic explanation at a cellular scale with variables that could be manipulated and observed. Its achievement is precise within its experimental system: it describes how electrically excitable membrane behaves under measured conditions. It does not, by itself, yield a theory of memory, visual categorization, or human judgment. A cross-level research program must add organization, synaptic connections, task conditions, and behavioral readouts. Hodgkin & Huxley (1952)

Theoretical neuroscience subsequently assembled mathematical descriptions of spiking variability, tuning, population codes, plasticity, information, and dynamics. Dayan and Abbott's synthesis systematized how models at different scales can express distinct scientific questions, ranging from neuronal responses to learning rules and networks. Models can be mechanistic when their variables closely correspond to cellular processes, or abstract when they summarize a population or computational function. The choice of abstraction should follow the target phenomenon and available measurements. At the circuit level, many parameters become latent because individual connections, synaptic efficacies, and internal states cannot be measured simultaneously in the same behaving preparation. Thus, an excellent cell-level model can coexist with substantial uncertainty about its behavioral embedding. Abbott & Dayan (2001) Hodgkin & Huxley (1952)

Two consequences follow for the interpretation of neural simulation. Reproducing a spike waveform validates a constrained account of excitability, whereas reproducing the choices of an animal validates a different set of predictions. A network constructed from plausible neuronal units can still implement an incorrect decision policy; a mathematically abstract model can accurately predict behavioral adaptation while making weak claims about synaptic realization. Cross-level validity is improved by using multiple measurements and by testing whether a common parameterization predicts both local dynamics and behavior. It is further improved by independent perturbations that were not used to fit the model. The demand for such out-of-fit prediction is a methodological consequence of the gap between explanatory scales. Abbott & Dayan (2001) Kriegeskorte & Douglas (2018)

The Historical Separation of Computational and Physiological Questions

Marr's framework reoriented vision research by asking what problem a perceptual system solves and what representations and transformations would make the solution possible. The 1976 Marr–Poggio report described formal levels before later cognitive computational neuroscience became a named research program. Its significance lies in the explicit separation of a computation's purpose from the particular biological means of performing it. Historical priority should be attached to the identifiable report and its claims, not projected backward from contemporary machine-learning practice. The framework encouraged investigators to state which measurements test task definitions, algorithms, and neural implementation. This is especially important when a computational model makes compelling behavioral predictions without a neuron-by-neuron correspondence. Marr & Poggio (1976)

Later systems neuroscience gained methods capable of sampling activity across many units and complex tasks. Multivariate functional imaging revealed distributed representational patterns, exemplified by Haxby and colleagues' 2001 study of category-dependent responses in ventral temporal cortex. The significance of that experiment was not simply that one category occupied one anatomical patch. Instead, distributed patterns included information even outside regions of maximal category response. This observation made the structure of population activity an object of theoretical explanation and altered the standards for relating cognitive categories to brain organization. Mapping moved from one structure–one function assignments toward tests of representations distributed across recorded channels. Haxby et al. (2001) Kriegeskorte et al. (2008)

The rise of predictive modeling introduced a corresponding risk of explanatory overstatement. Models trained to predict a brain signal or solve a cognitive task can be evaluated quantitatively, yet different architectures may achieve comparable predictions. Reviews of cognitive computational neuroscience therefore emphasize task competence, neural predictivity, and comparisons across conditions, whereas methodological critiques insist on rigorous behavioral descriptions. These perspectives belong to a common scientific conversation. The history is best read as a gradual expansion of complementary tests—from direct cellular mechanism, through computational decomposition, to population-level prediction and manipulation—rather than a replacement of one method by another. Kriegeskorte & Douglas (2018) Krakauer et al. (2017) Glaser et al. (2019)

Research Designs and the Logic of Joint Prediction

A cross-level study begins with an independently defined behavioral task and a transparent model of the participant's choices or responses. Experimental manipulations should distinguish what the model predicts from what its competitors predict. If reward volatility is varied, for example, a learning model should specify how the latent prediction error and updating parameter change on each trial; a neural hypothesis should separately specify which response component follows those changes, at what latency, and under what sensory or motor confounds. Parameters derived from behavior enter the neural analysis with uncertainty, and model comparison should include alternatives that can explain the same performance. The aim is a conjunction of discriminating predictions rather than two unrelated good fits. O'Doherty et al. (2007) Schultz et al. (1997)

An observation model is essential. Spiking measurements have millisecond-scale access to selected units, while BOLD measurements integrate neurovascular processes over slower times and larger populations. Logothetis and colleagues' simultaneous electrophysiology and imaging study provides an empirical reference for interpreting these modalities; measured signals cannot be treated as interchangeable direct samples of the same neural variable. Recording modality, spatial sampling, hemodynamic transform, temporal filtering, preprocessing, and measurement noise determine which properties a model can reasonably predict. A theory that matches voxel activity after flexible linear readout has met a different test from a theory that predicts spike timing without re-estimation. These tests should be described precisely in the report. Logothetis et al. (2001) Diedrichsen & Kriegeskorte (2017)

Good validation divides construction, fitting, model selection, and evaluation. When both neural and behavioral data influence the choice of architecture, the final neural match must be assessed on held-out observations with suitable independence. Strong designs also test transfer to new stimuli, task rules, subjects, or experimental interventions. Model parameters may be uncertain, correlated, or practically unrecoverable, so recovery simulations are important before treating estimated variables as psychological or biological quantities. A single fit statistic cannot discriminate a causal mechanism from a correlated proxy. Cross-level inference is most persuasive when constraints from different measurements converge and when a perturbation causes the specific joint change predicted by the theory. Glaser et al. (2019) Poldrack (2006) Kriegeskorte & Douglas (2018)

Measurement Models: Spikes, Field Potentials, and Hemodynamics

Electrical and imaging modalities observe different physical consequences of neural activity. Intracellular recordings can constrain membrane conductance models; single-unit and multiunit recordings sample spiking from selected cells; local field potentials summarize electrical activity from local populations; and BOLD fMRI records a vascular signal influenced by neuronal and metabolic processes. Each observation is filtered by electrodes, sensors, anatomy, sampling rate, and preprocessing. Simultaneous intracortical and fMRI observations in nonhuman primates demonstrated relationships among these signals while revealing why interpreting the BOLD response as a count of action potentials is inadequate. A cross-level mapping must therefore specify the physical measurement path before assigning psychological meaning to signal fluctuations. Logothetis et al. (2001) Hodgkin & Huxley (1952)

The temporal scale of a task imposes additional restrictions. A rapid decision may evolve within hundreds of milliseconds while an fMRI response peaks much later; averaging across trials may conceal changes in the order of neural states that matter for choice. Conversely, a time-varying neural measurement may capture detailed activity that bears little on the behavioral quantity the experiment was designed to explain. The analyst should align latent states with task events, model the sensor transformation, and predefine whether the target is trial-level variation, condition-level differences, or a population average. Those targets support different explanatory claims and require different statistical error structures. Mante et al. (2013) Logothetis et al. (2001)

Sampling also changes what absence of evidence means. Failure to record a hypothesized representation from one cortical site may reflect sparse sampling, a task-dependent code, nonlinear mixing, or genuinely absent activity. A positive decoding result establishes accessible information under a chosen training and testing procedure, yet may reflect shared sensory input or motor output rather than the causal computation under study. Evaluating a measurement model therefore involves spatial and temporal reliability, independent data partitions, appropriate nuisance regressors, sensitivity analyses, and explicit uncertainty. The neural signal is an empirical observable; the intervening psychological computation remains a hypothesis requiring stronger cross-level tests. Diedrichsen & Kriegeskorte (2017) Poldrack (2006)

Representational Similarity as a Bridge Between Brain and Model

Representational similarity analysis (RSA) compares the geometry of activity patterns across experimental conditions. For each pair of stimuli or task states, a dissimilarity is computed between the corresponding multichannel patterns; the resulting representational dissimilarity matrix (RDM) characterizes which conditions are separated and by how much. Kriegeskorte, Mur, and Bandettini introduced RSA as a practical hub linking brain recordings, behavioral judgments, and computational representations without requiring one-to-one identity between neurons, voxels, and artificial network units. A match means that the measured conditions show a comparable relational structure under the selected metric and sampling process. It is evidence for representational correspondence at a specified abstraction, not literal anatomical equality. Kriegeskorte et al. (2008)

The method is particularly suitable for comparisons that lack aligned units. Human fMRI measurements, primate electrophysiology, model activations, and behavioral similarity judgments can share a set of stimulus conditions while differing profoundly in physical realization. Comparing RDMs evaluates common representational structure, provided that stimuli, condition averaging, noise estimates, and sampling are properly controlled. The original work demonstrated applications in visual regions with models of differing complexity, and subsequent toolbox work specified procedures for model evaluation, inference, and visualization. Choosing dissimilarity measures is consequential: correlation distance, Euclidean distance, and cross-validated distances express different assumptions about noise and scaling. Kriegeskorte et al. (2008) Nili et al. (2014)

Representational agreement does not identify direction of information flow, synaptic pathway, or causal necessity. Two systems can have comparable category geometry because they were trained on similar inputs, optimized for the same task, or share broad statistical regularities. A strong RSA analysis thus compares plausible baseline models, controls representational reliability, uses independent data for model fitting and testing, and examines differences among competing geometries. The investigator should specify the model's layer or state and the brain region or latency being tested before interpreting results. Such care retains RSA's distinctive strength as a bridge of information geometry while preventing a resemblance in matrices from being promoted into a complete circuit explanation. Diedrichsen & Kriegeskorte (2017) Nili et al. (2014)

Encoding Models, Decoding Models, and Pattern Components

Encoding analysis predicts measured neural responses from candidate features or latent computational states. A researcher may estimate how image statistics, model-layer activations, or decision variables map onto voxel activity or firing rates. Its scientific meaning depends on which features were defined before fitting, the flexibility of the readout, and the held-out data used to assess prediction. A good encoding score indicates that a chosen representation contains information useful for predicting the measured response. It does not prove that the neural system explicitly computes the model's variables in the same way. Diedrichsen and Kriegeskorte placed encoding, pattern-component modeling, and RSA within a common framework of representational hypotheses. Diedrichsen & Kriegeskorte (2017)

Decoding reverses the statistical question by predicting stimulus identity, behavior, or task condition from neural measurements. Above-chance decoding shows that condition-related information is recoverable under the chosen classifier, cross-validation scheme, and data distribution. The existence of recoverable information should be interpreted separately from the claim that downstream neurons actually use this information to generate behavior. Classification may rely on correlated signatures outside the causal pathway. Similarly, reverse inference from a region's activation to a specific psychological process requires information about how often the same region activates in competing processes. Poldrack's analysis of reverse inference remains important for resisting intuitive but invalid one-region–one-function readings. Poldrack (2006) Haxby et al. (2001)

Pattern-component modeling explicitly represents the second-moment structure of activity profiles and can compare hypotheses about representational covariance. Its relation to RSA and encoding reveals why superficially similar analysis labels may test different likelihoods and summaries. A rigorous comparison declares assumptions about pattern noise, regularization, independence of conditions, and estimation of nuisance variables. Differences in the number of fitted parameters and the structure of training sets can make one method appear stronger for reasons unrelated to the scientific theory. A useful methodological report presents candidate representations, the measurement likelihood, the partition structure, and the quantity being generalized. With these elements, cross-level mapping becomes a reproducible hypothesis test rather than an interpretive analogy. Diedrichsen & Kriegeskorte (2017) Nili et al. (2014)

Model-Based fMRI and Trial-Wise Computational Variables

Model-based fMRI operationalizes a cross-level bridge by fitting a computational model to task behavior and testing whether its predicted latent variables covary with regional BOLD responses. O'Doherty, Hampton, and Kim described this strategy in reward learning and decision-making, where estimates such as expected value and prediction error can be generated on each trial. The procedure is scientifically informative because the task model specifies a time-varying regressor with computational meaning rather than treating each condition as a descriptive label. Yet a statistical association between the model regressor and BOLD signal remains an association mediated by a hemodynamic observation process. The identity of the underlying circuit operation must be tested with additional evidence. O'Doherty et al. (2007) Logothetis et al. (2001)

Behavioral fitting creates opportunities and hazards. The apparent neural correlate of prediction error may change when learning rates are fixed rather than individually estimated, when participants follow alternative strategies, or when regressors for outcome surprise and motor response overlap. Trial-wise latent variables often correlate with one another, reducing the capacity to discriminate value, salience, uncertainty, and updating accounts. The analysis therefore requires explicit alternative models, design diagnostics, sensitivity checks, and interpretable uncertainty about fitted parameters. Its inferential unit also matters: a group-level BOLD effect does not imply that a parameter is identically expressed in every participant's neural circuitry. Such distinctions are central to scientific model translation. O'Doherty et al. (2007) Poldrack (2006)

Cross-level evidence becomes stronger when behavior and imaging are jointly constrained rather than used sequentially without feedback. A hypothesized computational variable should predict neural activity beyond task timing and shared stimulus history, while neural data should disambiguate behavioral models that fit choices similarly. Prospective designs can test new reward schedules, perturbation effects, and individual differences without reselecting models after seeing outcomes. This approach also clarifies the interpretation of model failures: an inadequate behavioral model produces unreliable regressors, whereas an adequate behavioral model may still lack a faithful proposed neural realization. Both uncertainties should be reported rather than compressed into a single measure of fit. Kriegeskorte & Douglas (2018) Glaser et al. (2019)

Reward Prediction Error: A Canonical Cross-Level Case

Schultz, Dayan, and Montague's 1997 paper connected observations of dopaminergic responses with formal ideas about reward prediction. The account explained why neural responses associated with unexpected reward and predictive cues could be interpreted through the difference between expected and obtained outcomes. The scientific influence of this work lies in linking a quantitatively defined learning signal to measured neuronal activity, permitting time-specific and task-dependent tests of a theory. The term prediction error refers to a comparison internal to a particular model; its correspondence with dopamine activity must be established in the sampled task and circuit, rather than assumed as a universal identity for all dopamine signaling. Schultz et al. (1997)

The model-derived prediction error changes as learning proceeds, and its sign and timing are important. A cue can acquire predictive value and alter the expected outcome, so a neuron responding strongly at one stage of training may behave differently after expectations stabilize. A cross-level test should compare the predicted shift with the observed neural response trajectory, while recording the actual sequence of stimuli and rewards. This differs from simply demonstrating that the animal likes a reward or that a brain region activates when reward arrives. It relates a computationally specified learning variable to an experimentally structured neural phenomenon. Schultz et al. (1997) O'Doherty et al. (2007)

The reward literature also illustrates how the scope of a landmark result can expand beyond its original evidence. Diverse dopaminergic populations, task demands, and behavioral measures require attention to heterogeneity, and a computational variable may have multiple biological correlates. An analysis focused on this dossier asks what exactly has been validated: the behavioral update rule, the neural correlational signature, or a perturbationally supported role in learning. Competing models may predict different responses to changes in outcome probability, timing, salience, or movement. The defensible synthesis recognizes an influential explanatory bridge and preserves the need for task-specific causal tests and independent replication across contexts. Schultz et al. (1997) Krakauer et al. (2017)

Predictive Coding and Hierarchical Visual Inference

Predictive coding proposes that hierarchical systems can use descending expectations and ascending discrepancies to organize sensory processing. Rao and Ballard's 1999 visual-cortex model specified feedback signals carrying predictions and feedforward signals carrying residual error. Their computational simulations developed receptive-field-like properties and offered a possible account of some contextual effects in visual cortex. The historical result is a model-based explanatory proposal supported by simulations and comparison with known receptive-field phenomena; the simulation does not alone establish the architecture of every biological cortical feedback circuit. A cross-level evaluation must identify neural signatures that uniquely favor hierarchical error processing over competing recurrent or feedforward accounts. Rao & Ballard (1999)

Several operations are often grouped under the broad term prediction. One is stimulus anticipation based on learned regularity; another is a mismatch response to violated expectation; a third is probabilistic inference about causes under noise. These operations are mathematically and experimentally distinguishable, even when they occur together. Neural suppression following an expected stimulus, for instance, may also result from adaptation or attention. The researcher should model these alternatives, match their predictions to temporal and spatial neural measurements, and manipulate the conditions that separate them. A meaningful result specifies whether the model predicts activity magnitude, temporal dynamics, representational geometry, or behavior. Rao & Ballard (1999) Friston (2010)

Friston's free-energy perspective enlarged the theoretical discussion by proposing a broad framework in which perception, action, and learning are described through probabilistic inference and model-based regulation. Its generality is intellectually productive, but empirical informativeness depends on a concrete generative model and discriminating measurements. If a framework permits numerous implementation choices, finding one configuration that fits neural data does not identify a unique mechanism. For this dossier, predictive coding is therefore treated as a family of formal hypotheses with distinct commitments about signal direction, precision weighting, and learning. Strong evaluations specify the task, parameterization, data likelihood, and interventions before judging explanatory success. Friston (2010) Kriegeskorte & Douglas (2018)

Population Dynamics and Recurrent Computation

Neuron-by-neuron descriptions can obscure a computation distributed across many interacting cells. Population-dynamical approaches treat neural activity as a trajectory through a high-dimensional state space and ask whether recurrent interactions constrain how this trajectory evolves. Mante and colleagues analyzed macaque prefrontal population activity during a context-dependent decision task and used a recurrent model to interpret flexible selection of relevant sensory information. The experiments showed that complex individual-neuron responses could coexist with interpretable population dynamics. The study connected a behavioral requirement, a circuit-level state-space description, and model-generated trajectories, while leaving questions about the exact underlying synaptic architecture for further investigation. Mante et al. (2013)

Dynamic models make a different kind of correspondence claim from static feature encoding. A static model asks whether a neural pattern represents task-relevant information at a moment. A dynamical model predicts the sequence of states and how it changes after perturbation, differing initial conditions, or task-rule changes. Two systems may have similar average tuning yet different transition structures, which can generate different future behaviors. Conversely, low-dimensional trajectories can resemble one another even if high-dimensional implementation differs. Cross-level validation should therefore assess temporal prediction, robustness to noise, stability of latent states, and task-specific links to overt choices or movements. Mante et al. (2013) Churchland et al. (2012)

The concept of an attractor or latent computational state is explanatory only when its defining behavior is tested. A plotted trajectory may visually suggest a stable state, rotation, or decision boundary; geometry alone does not determine causality. The investigator must examine whether proposed dynamical parameters are identifiable from the sampled trials, whether alternative dimensionality reductions change the appearance, and whether model perturbations predict recoverable neural and behavioral responses. Neural population analyses can reveal stable regularities at a scale inaccessible to single-neuron inspection, but a mapping from those regularities to psychological processes remains contingent on well-controlled task manipulations. Churchland et al. (2012) Sussillo et al. (2015)

Motor Cortex as a Dynamical System

Motor neuroscience offers a stringent test of cross-level explanation because neural activity, movement, and muscles can be measured in coordinated temporal sequences. Churchland and colleagues' 2012 study of reaching investigated population dynamics and reported structured rotational features in cortical activity, motivating a dynamical account of motor preparation and execution. The data complicate explanations that attribute every unit's activity to a fixed movement parameter. Instead, the collective trajectory can provide an intermediate language between neural discharge and temporally structured action. The inference remains tied to the reaching tasks and recorded populations; extrapolation to every kind of movement requires additional evidence. Churchland et al. (2012)

Sussillo and colleagues then trained recurrent neural networks to produce the muscle activity of reaching monkeys, examining whether task-appropriate network dynamics could reproduce features seen in motor-cortical measurements. Such modeling relates an output requirement to internally organized activity, yielding a constructive demonstration that a particular class of recurrent system can generate suitable commands. A constructive model supports possibility and quantitative prediction. It does not establish that biological synapses have the same trained weights or that the modeled optimization reproduces the organism's developmental history. Validation improves when novel reach conditions or causal neural manipulations discriminate the dynamical account from alternative mappings. Sussillo et al. (2015) Churchland et al. (2012)

Muscle activity is itself an intermediate variable between cortical commands and observed movement. External forces, limb biomechanics, sensory feedback, and control strategies all affect the trajectory of the hand. A model fitted exclusively to final movement may therefore conceal differences in the intervening processes, while a model fitted to muscles can still miss cortical mechanisms. An integrated study compares neural states, muscle activation, and kinematics under changes in load, target, context, and feedback. The evaluation should ask whether a common parameterization explains transitions at each scale. This illustrates how adding levels increases explanatory constraints while also increasing the need for correctly specified observation models. Sussillo et al. (2015) Abbott & Dayan (2001)

Visual Recognition and Task-Optimized Neural Networks

Object recognition is a productive setting for comparing trained artificial networks with biological sensory pathways. Yamins and colleagues trained hierarchical models for performance on visual object tasks and reported that high-performing networks predicted aspects of inferior temporal neuronal responses, with intermediate layers associated with responses in area V4. Their result joined task competence with neural predictivity under a defined model class. The claim concerns quantitative correspondences observed in the selected images, recording sites, and evaluation protocol. It does not entail that the primate brain uses the network's exact layers, training algorithm, artificial neuron nonlinearities, or connectivity pattern. Yamins et al. (2014)

A suitable cross-level comparison distinguishes the task objective, training distribution, architecture, readout flexibility, and test stimulus set. When a large network is optimized on category labels, its representations may encode natural-image statistics that also organize biological vision. The relationship can arise through convergent constraints imposed by the task, without copying biological circuitry. RSA provides one test of whether two systems preserve similar relative distances among images, while encoding analysis asks whether model features predict measured neural channels. These are related and nonidentical tests. The strength of the evidence depends on whether image selection and model tuning are independent of final neural evaluation. Yamins et al. (2014) Kriegeskorte et al. (2008) Diedrichsen & Kriegeskorte (2017)

Artificial neural-network models gain scientific value from being executable, perturbable, and capable of predicting new conditions. Their failures can be equally informative: sensitivity to stimulus transformations, differences in response timing, or discrepancies under altered tasks can reveal which constraints a proposed visual computation omits. Saxe, Nelli, and Summerfield emphasized the need to ask what scientific question a deep-learning model actually answers. Behavioral accuracy, neural predictivity, and biological plausibility must therefore remain separately documented outcomes. Treating them independently prevents the benchmark from substituting for an explanation of how a circuit produces visual behavior. Saxe et al. (2021) Glaser et al. (2019)

Multiple Cognitive Tasks and Flexible Neural Organization

Human and animal cognition requires transfer among tasks whose rules and sensory demands change. Yang and colleagues trained recurrent networks on 20 cognitive tasks and examined the organization of the resulting representations, identifying functional clusters in artificial units and structured relations among tasks. Their work illustrates how task diversity can provide a stronger test of representational organization than performance on one carefully selected benchmark. It also permits hypotheses about shared processes such as accumulation, maintenance, categorization, and inhibitory control. The result concerns an artificial model's learned organization and the scientific possibilities it creates; it does not directly establish a one-to-one account of any particular human cortical network. Yang et al. (2019)

Multitask models invite cross-level questions about compositionality. Do tasks that share behavioral computations recruit related neural subspaces? Does changing a rule reorganize population activity while preserving a stable component? Does a model trained on several tasks predict behavior on unseen combinations, or does it simply memorize the training menu? These questions require operational definitions of task similarity and independent behavioral and neural measurements. The task geometry derived from model units must be evaluated against recorded brain geometry and against alternative explanations based on shared stimuli and response mappings. A common label like working memory does not, by itself, establish equivalent representations across species or machines. Yang et al. (2019) Diedrichsen & Kriegeskorte (2017)

The comparison becomes especially informative when investigators vary the training distribution and architectural constraints of a computational model. A representational pattern that persists across multiple architectures and tasks may indicate a broad computational requirement, while one appearing only in a narrow training regime may reflect contingent optimization. Even robust cross-network patterns need neural and behavioral evidence before being characterized as biological mechanisms. The neuroscience-inspired AI literature identifies a reciprocal research program in which biological findings inform computational design and artificial models generate hypotheses for biology. The central scientific discipline is to distinguish hypothesis generation, task-level analogy, cross-level prediction, and interventionally supported explanation. Hassabis et al. (2017) Saxe et al. (2021)

Neural Variability, Noise Correlations, and Reliable Representation

Repeated presentations of nominally identical stimuli do not reproduce identical neural responses. Spike counts, timing, local field potentials, and population trajectories vary within and across trials, and experimental recordings include uncertainty from sensing and preprocessing. A representational comparison that uses only averages can therefore mistake measurement instability for a property of the computational code. Neural noise matters twice: it constrains how much information a measured pattern can carry, and it affects how accurately a model's predicted geometry can be compared with recorded patterns. Reliability must be measured at the same granularity as the claim, including the individual stimulus, task condition, and participant or animal. Diedrichsen & Kriegeskorte (2017) Nili et al. (2014)

Correlated variability can alter what a population-level decoder or encoder recovers. When recorded channels share noise because of common input or experimental state, standard independence assumptions fail; when different samples of neurons have different covariance structures, comparison across populations becomes sensitive to sampling. A strong analysis estimates noise covariance when data permit, uses cross-validation partitions that respect repeated observations, and provides uncertainty around estimated similarity or prediction. Cross-validated representational distances are especially useful because naive estimates of distances may be positively biased by noise. The researcher must still state which measurement model and covariance assumptions support the comparison. Diedrichsen & Kriegeskorte (2017) Kriegeskorte et al. (2008)

Variability is not automatically a defect to remove. Changes in arousal, attention, learning stage, or internal policy can generate real behavioral and neural variation; indiscriminate averaging may erase the mechanism being studied. Separating measurement noise from meaningful state variation requires independent behavioral indicators and, where feasible, repeated measurements that cover the proposed timescale of state change. A theory of flexible decision making, for example, should explain both stable task-specific representations and changes in those representations after a rule switch. Noise-sensitive model validation thus connects statistics with the psychological explanation rather than treating uncertainty as an afterthought. Mante et al. (2013) Krakauer et al. (2017)

Temporal Alignment, Neural Latency, and the Shape of a Computation

Cross-level models often have a hidden temporal assumption: a computational variable is assigned to a neural observation at the moment when the theory expects it to occur. In a trial involving sensory evidence, valuation, action selection, and feedback, multiple related variables may unfold in a partially overlapping sequence. A neural response recorded during stimulus onset may reflect early sensory drive; activity nearer response commitment may reflect accumulated information or motor preparation. When a model's latent variable is regressed on time-averaged activity, this distinction can disappear. Precise hypotheses therefore specify temporal windows, lags, event boundaries, and the observation process linking modeled and recorded time. Mante et al. (2013) Churchland et al. (2012)

The appropriate temporal measurement depends on the scientific question. Extracellular electrophysiology offers high-resolution neural dynamics in limited sampled populations. Hemodynamic imaging supports spatially distributed observations while requiring an explicit account of delay and temporal smoothing. These capabilities permit different tests: fast changes after a transient perturbation may be visible electrophysiologically, whereas slower imaging data can assess regional associations at a coarser timescale. Interpreting one modality with the timescale of another can produce spurious claims of sequential processing. Cross-modal studies should articulate what transforms a candidate neural process into each measurement. Logothetis et al. (2001) Glaser et al. (2019)

Computational neural dynamics are scientifically valuable because they constrain transitions rather than isolated states. A recurrent network may match neural activity at two moments while following an incorrect trajectory between them, or reproduce a population trajectory while generating the wrong behavioral latency distribution. The requirement for joint prediction therefore includes timing: when information emerges, how long it persists, what changes after a cue, and how that change predicts a response. Experiments varying stimulus duration, decision deadlines, or feedback timing can distinguish models that are similar under static tests. Such temporal manipulations make the bridge hypothesis discriminating and biologically interpretable. Mante et al. (2013) Sussillo et al. (2015)

Latent Dimensionality and the Geometry of Neural State Spaces

A neural population with thousands of recorded channels can exhibit activity that varies predominantly along a smaller number of dimensions. Dimensionality reduction constructs coordinates that summarize dominant patterns, making trajectories and task-dependent changes visually and statistically tractable. Yet a recovered axis or plane is defined by the data sample and algorithm, and it need not coincide with one anatomical pathway or psychological process. Recurrent-circuit interpretations become scientifically stronger when a model predicts the geometry of neural transitions in independently sampled conditions and links those trajectories to behavior. Churchland et al. (2012) Mante et al. (2013)

Representational geometry and dynamical geometry address different questions. An RDM reports the relative positions of condition-specific activity patterns; state-space analysis can describe how a population moves between those positions as the task unfolds. Two recordings may exhibit similar category geometry despite different response times, transition rates, or recurrent stability. Conversely, similar low-dimensional rotations can arise in distinct networks with different tuning at the individual-unit level. The validity of a cross-level comparison depends on which aspects of geometry are defined as equivalent and whether arbitrary rotations, rescalings, or flexible readouts are allowed. An apparently strong match can be weakened by a transformation class so unconstrained that many alternative models would qualify. Kriegeskorte et al. (2008) Diedrichsen & Kriegeskorte (2017)

Dimensionality estimates may also change with numbers of units, trials, and experimental conditions. A scientific report should therefore distinguish latent dimensions needed for predictive accuracy, dimensions inferred from neural noise structure, and dimensions to which a behavioral theory assigns meaning. Holding out task conditions can reveal whether a latent space supports generalization, while controlled perturbation can test whether a trajectory's proposed directions affect behavior. A model need not reproduce every recorded channel to be informative, but it should predict features that are stable, measurable, and theoretically consequential. Nili et al. (2014) Sussillo et al. (2015)

Model Readouts, Training Objectives, and the Risk of Flexible Correspondence

Neural-network models are often compared with brain activity by fitting a linear or nonlinear mapping from artificial units to biological channels. Such a readout can be legitimate: biological measurements and model units have different scales and do not need to correspond one-to-one. But additional flexibility expands the range of neural patterns a model can predict, creating a risk that benchmark performance reflects the readout rather than the proposed computation. The number of adjustable parameters, the degree of regularization, and the independence of model selection and final testing must therefore be disclosed. An interpretable comparison specifies the transformations permitted between representations and justifies their relation to the scientific question. Diedrichsen & Kriegeskorte (2017) Yamins et al. (2014)

Training objective is a second important source of interpretive variation. A network trained on object classification has acquired representations shaped by its training examples and loss function; a recurrent model trained to control motor output acquires dynamics shaped by output requirements; a multitask network's organization reflects the set of rules, targets, and feedback conditions used for training. Similarity to a neural recording can indicate that the engineered task imposes comparable functional constraints, but optimization history and biological implementation may differ substantially. The correct inference is a restricted constructive account: this architecture can solve the specified task and may predict particular neural measurements. Biological mechanism requires further evidence. Yamins et al. (2014) Sussillo et al. (2015) Yang et al. (2019)

Strong comparisons deliberately vary the readout, architecture, task objective, and training dataset while preserving a shared independent test. If a prediction depends on a highly flexible fitted mapping, the result should be compared with equally flexible alternative representations, including simpler baselines. If correspondence survives a restricted readout and transfers across stimulus families or task demands, the model acquires more specific support. Analyses should also report the effect of architectural changes and what aspects of observed neural dynamics remain unexplained. These tests transform the use of artificial networks from resemblance-based storytelling into a controlled research program. Saxe et al. (2021) Glaser et al. (2019)

Within-Subject, Between-Subject, and Cross-Species Mappings

An explanatory mapping can be constructed at different populations of inference. A within-subject model asks whether one participant's task behavior covaries with their neural measurements; a population model asks which relationships replicate across individuals; a cross-species model asks whether some task functions and neural characteristics share a meaningful structure across organisms. These are not equivalent predictions. A result recovered from one trained macaque during a carefully practiced decision task cannot automatically be treated as a universal property of human cognition, even when an analogous human brain region can be identified. Generalization requires matched tasks, justified measurement comparability, and explicit analysis of individual variability. Mante et al. (2013) Krakauer et al. (2017)

Cross-species comparisons can be powerful when precise neural signals are accessible in animal preparations and behavioral or imaging measurements are available in humans. RSA was specifically developed to enable cross-method comparisons by using condition-level representational geometry. Its promise depends on whether the compared stimuli and task demands induce sufficiently similar computations, and whether differences in sensory apparatus or learning history change the representational problem. Correspondence across species should be stated at the tested representational, computational, or behavioral level. A shared matrix of dissimilarities does not prove identical neurons, circuits, or conscious experiences. Kriegeskorte et al. (2008) Nili et al. (2014)

Hierarchical and multilevel analyses can account for genuine variation rather than removing it by pooling. Model evaluation should specify how parameters are shared across individuals, what sample size supports claims about between-person differences, and whether measurement quality is comparable across recording sessions and populations. Training history may be especially consequential: artificial systems can be optimized on massive datasets, whereas biological subjects accumulate experience through development and bodily action. Cross-system and cross-species research should therefore include systematic comparisons of the learning environments that shaped observed representations. This keeps the Artificial Era comparison empirically grounded and indicates precisely where a proposed mapping is robust. Hassabis et al. (2017) Saxe et al. (2021)

Connectomics and the Structure–Function Bridge: New Evidence from 2024–2025

Contemporary connectomics moves cross-level mapping closer to the physical architecture of circuits. Dorkenwald and colleagues' 2024 whole-adult-brain reconstruction in Drosophila reported approximately 139,255 neurons and 50 million chemical synapses, with cell and connection annotations intended to support systematic modeling. This extraordinary anatomical resolution permits hypotheses about how identified pathways could route information through the brain of one species. Yet a wiring diagram records connectivity, not the full time-varying state of its synapses or the neural signals produced in a particular behavioral episode. Transforming anatomy into a psychological explanation requires measured activity, operationalized behavior, and a dynamic model specifying which connections contribute to the tested computation. Dorkenwald et al. (2024) Hodgkin & Huxley (1952)

The MICrONS Consortium's 2025 investigation linked dense two-photon calcium imaging of approximately 75,000 visual-cortical neurons with electron-microscopy reconstruction of more than 200,000 cells and approximately half a billion synapses in a mouse cortical volume. Co-registration of physiological responses and anatomical reconstruction supplied an unusually direct opportunity to compare visual function with physical connectivity. The quantitative scale is impressive, but the source explicitly reports a dataset derived from one mouse, a particular volume, experimental stimuli, and extensive reconstruction and registration procedures. Its evidentiary contribution is a major enabling resource for model testing at the structure–function interface, and its sampling boundaries must remain visible in generalizations. MICrONS Consortium (2025)

The major methodological advance is that functional response profiles and synaptic anatomy can be evaluated within overlapping sampled tissue rather than inferred from separate experimental populations. Even this integration leaves gaps between observed connections, causal influence, whole-organism behavior, and psychological constructs. Calcium indicators, optical sampling, segmentation, registration uncertainty, and functional selectivity affect the reliability of the proposed mapping. A network-derived prediction must be tested under new visual conditions, and claims about causal routing would benefit from perturbation and independent preparations. The scientific direction is toward joint measurement of anatomical structure, neural dynamics, and behavior, with constraints that are explicitly traceable to each measurement process. MICrONS Consortium (2025) Logothetis et al. (2001) Krakauer et al. (2017)

Connectivity Rules, Functional Selectivity, and Mechanistic Prediction

A structure–function correspondence is more precise when it specifies a quantitative wiring rule rather than merely noting that connected neurons are both active. In a companion 2025 MICrONS study, Ding and colleagues examined the relationship between functional response similarity and synaptic connectivity among mouse visual-cortical neurons across layers and areas. Their findings supported a tendency for neurons with similar measured response properties to be preferentially connected, including within and across visual areas. The observation constrains candidate connectivity models because a plausible mechanism must account for systematic relationships between tuning and the probability or structure of synaptic connections. It does not establish that functional similarity causes the connections, or that every connection follows the tendency. Ding et al. (2025)

The experiment illustrates a central challenge of cross-level inference: measured similarity can arise from shared input, developmental history, common task exposure, or selective wiring, while the reconstruction provides direct anatomical evidence about one part of the causal system. A meaningful explanatory model compares these possibilities through specific predictions about connection statistics, feature tuning, circuit dynamics, and changes after controlled manipulations. Spatial proximity, cell type, layer, and sampling must be included when interpreting observed connectivity patterns. The anatomical result also has a distinct inferential scope from a theory of perceptual categorization or report, which requires an additional link to behavior. Ding et al. (2025) MICrONS Consortium (2025)

These datasets identify an emerging class of cross-level tests with fewer freely adjustable assumptions. An investigator can derive a circuit-level prediction from an observed connectivity graph, simulate neural responses to held-out stimuli, and compare those outputs with physiologically observed activity and behavioral measures. Success would provide multiple constraints at the same biological scale, while failure could reveal missing synaptic properties, input signals, recurrent dynamics, or an incorrect computational task description. Functional connectomics therefore strengthens the empirical basis for mechanistic modeling and simultaneously increases its responsibility for precise measurement, causal logic, and uncertainty. The 2024–2025 results are scientific advances in specific preparations, forming a research platform for wider generalization. MICrONS Consortium (2025) Ding et al. (2025) Dorkenwald et al. (2024)

Causal Interventions Across Levels

A causal neural explanation requires an intervention whose consequences discriminate among proposed mechanisms. If an investigator claims that a neuronal population integrates evidence to control a decision, the hypothesis should specify how temporarily changing its state alters subsequent neural dynamics and choices, including expected direction, timing, and task dependence. Selective perturbation, pharmacological manipulation, lesions, or stimulation may provide evidence of necessity or contribution, but their selectivity and compensatory effects require analysis. A change in behavior after broad neural disruption establishes involvement more readily than it establishes the precise latent computation. The inferential target is a predicted causal pattern linking task manipulation, neural state, and measured behavior. Krakauer et al. (2017) Mante et al. (2013)

The distinction between necessity, sufficiency, and mediation is especially important. Silencing a region that participates in many functions may impair performance without showing that its measured activity represents the model's chosen variable. Artificial activation may elicit a response under laboratory conditions without proving the same activation normally causes that response. To test mediation, the researcher needs a mechanism specifying how the perturbation changes an intermediate computation and why that intermediate change produces a downstream behavior. Such tests can be triangulated with recordings and computational simulations, but simulation interventions alone are not biological interventions. Mechanistic claims should be limited to the tested intervention, observed effect, and justified inferential assumptions. Krakauer et al. (2017) Kriegeskorte & Douglas (2018)

Perturbations also provide unusually informative model-selection opportunities. Two models that agree on unperturbed choice probabilities may diverge when sensory evidence is degraded, a population state is briefly displaced, or the available action set changes. Intervention design should target such differences and preregister the expected pattern when feasible. A well-supported bridge includes an observed neural trajectory, a computational account that predicts its response to perturbation, and a consequent behavioral change with specified timing. Failure of the predicted cascade may localize the mistake to the behavioral algorithm, neural mapping, or observation process. The value of causal experiments is therefore their capacity to expose where an explanatory connection fails. Mante et al. (2013) Churchland et al. (2012)

Identifiability, Observational Equivalence, and Model Recovery

Different computational architectures can generate indistinguishable output distributions under a limited set of tasks. A fitted learning rate, time constant, or representational dimension may therefore lack a unique psychological interpretation even when the model predicts behavior accurately. Structural identifiability concerns whether the model's parameterization permits unique inference from ideal data; practical identifiability concerns whether finite, noisy measurements can distinguish relevant values. Cross-level modeling adds the possibility that several neural implementations realize the same algorithm and that several behavioral algorithms produce similar population statistics. Evidence for one correspondence must be judged against this space of alternatives. Diedrichsen & Kriegeskorte (2017) Glaser et al. (2019)

Simulation-based recovery provides a direct diagnostic. Parameters or model identities are used to generate synthetic observations under the planned task and noise process; the original estimation pipeline then attempts to recover them. A procedure that cannot reliably distinguish the simulated alternatives is unlikely to adjudicate them from the corresponding experimental dataset. Recovery should include realistic trial counts, subject variability, nuisance effects, and signal-to-noise conditions. The analysis should also inspect parameter tradeoffs, sensitivity to priors and regularization, and whether allowing a flexible readout makes otherwise distinct representational models observationally similar. Such diagnostics do not guarantee truth, but they clarify which claims the experimental design is able to support. Diedrichsen & Kriegeskorte (2017) Nili et al. (2014)

Cross-level data may improve identification when they supply genuinely independent constraints. A computational model predicting identical choices for two parameter sets may predict different neural time courses, while different models that fit averaged fMRI signals may diverge in reaction times or electrophysiological latencies. The advantage disappears when the additional measurement is merely another transformed expression of the same nuisance variable. Model comparison should therefore examine incremental predictive information and the causal logic of each channel, not assume that a greater number of signals automatically means stronger explanation. A carefully delimited unidentified model is preferable to a richly narrated but empirically equivalent mechanistic story. O'Doherty et al. (2007) Logothetis et al. (2001)

Statistical Validation and Generalization Beyond the Fitting Set

Generalization tests whether a model has learned stable relationships rather than accidental features of a specific dataset. In neural prediction, held-out observations can be separated by trials, stimuli, participants, recording sessions, or task conditions; these choices answer different scientific questions. Randomly dividing nearly identical trials may inflate apparent generalization when shared temporal or stimulus structure crosses the partition. Training and testing should reflect the population and circumstances for which the theory makes claims. If preprocessing or model selection uses information from the test set, the resulting accuracy no longer represents an independent evaluation. The same principle applies to behaviorally fitted latent variables later used as neural predictors. Glaser et al. (2019) Yamins et al. (2014)

A neural benchmark should also include a principled ceiling and useful baselines. Measurement noise, biological variation, sampling reliability, and the temporal stability of a representation place limits on attainable model correspondence. Comparing an elaborate network only with an implausibly weak baseline can make its fit appear decisive even when simpler features or alternative architectures predict nearly as well. RSA toolboxes and representational-model analyses provide procedures for noise-sensitive comparison, but the precise inferential assumptions must be documented. Strong evaluations report uncertainty around model differences, variability among subjects, stability across data partitions, and results for genuinely competing explanations. Nili et al. (2014) Diedrichsen & Kriegeskorte (2017)

Transfer tests are most revealing when they challenge the proposed mechanism. A visual model can be evaluated under changed viewpoint, unusual texture, or new categories; a learning model under reversal or altered contingencies; a cognitive-control model under different rule combinations. Such shifts distinguish task-general computational constraints from dataset-specific shortcuts. A successful transfer establishes a wider predictive range under the tested conditions, while failure identifies a boundary of the model. Neither outcome should be interpreted as a direct certificate of biological realism. The scientific aim is an accumulating sequence of risky predictions about behavior and neural states, where model restrictions are refined as the evidence changes. Yamins et al. (2014) Yang et al. (2019)

Reverse Inference and the Interpretation of Neural Correlates

Neural correlates acquire psychological meaning through the design that elicits them. If a brain region becomes active during a task described as attention, it does not follow that the region exclusively implements attention or that every activation indicates attention. Poldrack's analysis of reverse inference emphasized the logical difficulty of inferring a cognitive process from activation without considering the frequency of activation across alternative processes. Computational models improve this situation by producing explicit trial-level or state-level predictions, but a fitted neural correlate remains vulnerable to omitted variables and correlated task structure. The reader must ask which observations were expected specifically under the proposed model and which would also arise under rivals. Poldrack (2006)

A particularly common ambiguity occurs when reward, salience, task difficulty, reaction time, surprise, and movement vary together. A model-based neural regressor may correlate with one of these quantities even if its intended psychological process is not the primary cause of the recorded response. Orthogonalized task manipulations, carefully specified nuisance models, temporal dissociation, and replication across contexts can help resolve these cases. Yet no single statistical adjustment can rescue a design in which the critical theoretical variables never vary independently. Cross-level mapping requires the experimental manipulation to create discriminating variation at the level of both behavior and neural observation. O'Doherty et al. (2007) Poldrack (2006)

Positive results should be described in a hierarchy of claims. A response differs across task conditions; a decoder recovers condition identity; an encoding model predicts recorded activity; a computational variable explains additional variance; a perturbation changes the predicted circuit state and behavior. These propositions offer progressively different evidence rather than interchangeable synonyms for mechanism. They also reveal how a complex theory may succeed in one respect and fail in another. In reporting a model, it is scientifically productive to specify the strongest supported claim at each level and the next experiment that could strengthen or overturn the bridge. This discipline advances explanatory clarity across methodologies. Diedrichsen & Kriegeskorte (2017) Krakauer et al. (2017)

Behavioral Decomposition as an Independent Scientific Constraint

A neural explanation becomes ambiguous when behavior is reduced to a single accuracy score. The same percentage of correct responses can result from different decision thresholds, evidence quality, lapses, learning histories, or policies. Behavioral decomposition studies the timing, errors, conditional dependencies, and adaptation of actions to distinguish these possibilities. Krakauer and colleagues argued that careful behavioral analysis is central to understanding the neural implementation of cognition. Computational neuroscience should therefore specify the behavioral process before treating an observed circuit signature as its physical instantiation. A well-designed task is an instrument for distinguishing algorithms, not merely a way to induce brain activation. Krakauer et al. (2017)

Trial-by-trial measurements can reveal dissociations hidden in condition averages. Choice reversals, response latency, confidence reports, trajectory corrections, and sequential dependencies provide different constraints on a cognitive model, although each demands its own measurement account. Linking these quantities to neural population trajectories can distinguish a sensory encoding account from a motor preparation account or a dynamical-control hypothesis. Behavioral sampling must also account for training stage, fatigue, motivational state, and strategies that change during the experiment. The best bridge model explains the observed temporal organization of action in a manner consistent with the recorded neural signal, rather than retrofitting a verbal interpretation to a correlational map. Mante et al. (2013) Churchland et al. (2012)

Cross-level work is strengthened by comparative experiments within and across individuals. A behavioral mechanism inferred from one trained animal or one narrow laboratory task may have uncertain relevance to a broader population. Model parameters should be interpreted relative to the tasks and measurements used to identify them, and heterogeneous strategies may warrant mixtures or hierarchical descriptions rather than a universal individual model. Such considerations also shape clinical and translational interpretation, where neural measures often change with age, context, and measurement hardware. A robust explanation describes where the mapping holds, what population it covers, and how competing behavioral decompositions were evaluated. Krakauer et al. (2017) Glaser et al. (2019)

Evaluation of the Evidence: Strengths and Limits of Landmark Studies

The evidence reviewed in this monograph spans several fundamentally different study designs. Hodgkin and Huxley fit a quantitatively explicit account of membrane currents under controlled physiological conditions; Schultz and colleagues related patterned dopaminergic responses to a temporal prediction model; Haxby and colleagues characterized distributed category information in human imaging; and Mante and colleagues interpreted primate decision-related population dynamics with recurrent computational models. The studies offer distinct constraints on an explanatory bridge. Their precision comes from operational control and explicit variables, while their scope remains bounded by recording modality, preparation, task, sample, and analytical assumptions. Aggregating their conclusions into one global measure of explanatory accuracy would erase the very differences that make them instructive. Hodgkin & Huxley (1952) Schultz et al. (1997) Haxby et al. (2001) Mante et al. (2013)

At a second level, methodological research establishes how to compare proposed mappings. RSA offers geometry-based comparison across nonmatching channels; pattern-component and encoding analyses test different aspects of representational models; model-based fMRI makes cognitive variables explicit; and supervised learning provides flexible predictive benchmarks. These techniques contribute standards of quantification, reproducibility, and discrimination, but each has possible ambiguities. Model fit is vulnerable to overparameterization, decoding may exploit correlations, imaging depends on its observation process, and behavioral latent variables can be poorly identified. A credible synthesis must retain the specific strengths and inferential limits of each method rather than assign all computational evidence the same status. Kriegeskorte et al. (2008) Diedrichsen & Kriegeskorte (2017) O'Doherty et al. (2007) Glaser et al. (2019)

Systematic quantitative effect comparisons would require common outcomes, explicit search and inclusion criteria, comparable study populations, and standardized assessment of bias, none of which is present across these diverse model families. This research therefore does not pool published coefficients or rank their quantitative strength. It provides a structured critical interpretation: physical measurements constrain mechanisms; multivariate comparisons constrain representational structure; model-based tasks constrain proposed latent computations; and interventions constrain causal pathways. Where multiple independent constraints agree on a specific prediction, confidence in that bridge can increase. Where only an abstract resemblance or trained predictive fit is available, the inference remains representational or functional. Kriegeskorte & Douglas (2018) Krakauer et al. (2017)

Artificial Models and the Artificial Era: Scientific Scope

Artificial networks contribute to cross-level neuroscience because they supply explicit, executable hypotheses about how computations might be organized. Trained systems permit targeted ablation, modification of architecture, controlled training histories, and measurement of every internal unit. This makes them useful test beds for theoretical claims about hierarchy, recurrence, distributed representation, and task optimization. Work on neuroscience-inspired AI documents this exchange, while task-optimized vision and multitask recurrent models demonstrate concrete methods for proposing neural and behavioral tests. Their methodological value rests on the experiment they make possible: a specified model predicts independently measured characteristics of biological cognition under conditions not chosen solely to favor the model. Hassabis et al. (2017) Yamins et al. (2014) Yang et al. (2019)

An artificial network is also a distinct engineered system with its own material implementation, training process, and observation opportunities. The same mathematical computation can be realized in silicon and biological tissue through different causal organizations. Similarity in task performance or representational geometry can justify comparative functional hypotheses, while claims of architectural equivalence require closer tests of dynamical and intervention responses. Interpreting artificial models as scientific instruments preserves their explanatory potential without assuming that optimizing a model for a behavioral benchmark reproduces the evolutionary, developmental, or biological constraints of a nervous system. The distinction is particularly salient when a large network has enough flexibility to fit many different task and neural outcomes. Saxe et al. (2021) Glaser et al. (2019)

Within the Artificial Era, cross-level modeling also raises questions about what comparisons between Homo and Artificial can scientifically establish. Behavioral, computational, representational, and physical comparisons should be operationalized separately. A machine's success on a task can demonstrate an engineered capability; a shared latent geometry can demonstrate structural similarity under an analysis; an intervention can reveal how the artificial system implements its own computation. None of these findings alone resolves questions about psychological identity, subjective experience, or the biological instantiation of a human mental process. The productive scientific program is to create matched tests with transparent model histories, controlled environments, and explicit claims about the levels at which correspondence holds. Hassabis et al. (2017) Kriegeskorte & Douglas (2018)

Relations to Psychology, Cognitive Neuroscience, and Clinical Translation

Cross-level neuroscience belongs within psychology because the psychological task and construct define what needs explaining. It is also a biological research enterprise because neurons, synapses, circuits, and measured tissue responses constrain implementations. Mathematics contributes formal models, statistics quantifies uncertainty, and computer science supplies constructive systems for testing possible algorithms. The interdisciplinary relationship is therefore precise: behavioral experiments identify explanatory targets, computational theories propose generative rules, and neuroscientific measurements and interventions test biological realization. Treating one discipline's vocabulary as a substitute for another's measurements would undermine this structure. The joint program is strongest when each theory states which level it addresses and which evidence could disconfirm its proposed bridge. Kriegeskorte & Douglas (2018) Krakauer et al. (2017)

Translational uses require further safeguards. An estimated learning parameter or imaging correlate may be useful for identifying behavioral heterogeneity or describing a research cohort, but its biological interpretation depends on validity in the population where it will be used. A model of healthy volunteers under laboratory reward schedules does not automatically capture psychiatric symptoms, developmental changes, or medication effects. Translational claims demand population-specific calibration, measurement reliability, independent predictive testing, and comparison with existing behavioral or clinical assessments. This monograph is concerned with the scientific logic of translation rather than with treatment recommendations or individual diagnosis. The distinction protects the evidentiary relationship between neural computation and psychological application. O'Doherty et al. (2007) Poldrack (2006)

The same framework clarifies neighboring research dossiers. Formal Psychological Models and Identifiability addresses model recovery in general; Reinforcement Learning and Model-Based Control studies learning and policy mechanisms; Computational Psychiatry and Clinical Model Translation examines clinical applications; Machine Learning Models as Scientific Models of Human Cognition evaluates artificial models as cognitive theory. The present dossier occupies the interface linking behavioral latent states and biological mechanisms, asking which observations and interventions support cross-scale correspondence. These boundaries permit focused collaboration among related fields while preserving a unique research object. They also prevent the general scientific term computational neuroscience from becoming a container for every research problem involving algorithms or neural recordings. Diedrichsen & Kriegeskorte (2017) Saxe et al. (2021)

Open Problems and a Cumulative Research Agenda

Several open problems determine progress in cross-level explanation. First, observation models are incomplete: measured electrophysiology and imaging provide partial, differently filtered access to neuronal processes. Second, model equivalence remains widespread because many architectures can predict a narrow dataset. Third, task descriptions may be too coarse to identify the behaviorally relevant state or strategy. Fourth, neural recording often covers only a subset of the circuit while causal interventions affect broad or compensating systems. These problems interact. Better data alone will not repair an ambiguous task model, and a more sophisticated algorithm alone will not resolve an inadequately characterized measurement process. Their joint diagnosis should guide the design of new studies. Logothetis et al. (2001) Krakauer et al. (2017) Glaser et al. (2019)

A cumulative agenda requires datasets that couple rich behavioral measurements to neural responses under prospective model tests. Researchers can predefine representational and dynamical hypotheses, publish the analysis pipeline, report parameter recovery, benchmark strong alternatives, and reserve conditions or individuals for independent evaluation. Where ethically and technically feasible, interventions should be chosen because candidate models predict different effects. Reports should include failures and uncertainty, not only successful correlations. Cross-laboratory comparisons can then establish whether a mapping survives changes in equipment, task implementation, analysis choice, and population. Such evidence would reveal whether apparent neural–computational alignment reflects a portable mechanism, a robust abstract regularity, or an artifact of a specific benchmark. Nili et al. (2014) Diedrichsen & Kriegeskorte (2017) Kriegeskorte & Douglas (2018)

Theoretical research can further improve bridge hypotheses by specifying what counts as the same computation under different implementations. One candidate criterion is predictive equivalence across a defined class of input perturbations; another is similarity in a dynamical state transition; another is representational geometry under controlled stimuli. Each criterion has a different scientific scope and should be tested accordingly. Future models should explicitly relate these criteria instead of treating them as interchangeable. Comparison with artificial systems adds opportunities for controlled internal intervention and alternative architectures, while human and animal experiments establish which correspondences hold biologically. The resulting program organizes scientific uncertainty into experimentally answerable questions. Yang et al. (2019) Hassabis et al. (2017) Saxe et al. (2021)

Conclusion

Computational neuroscience explains psychological functions more fully when it connects the mathematical structure of a task to an empirically constrained account of nervous-system activity. The history runs from quantitative membrane biophysics and computational-level analysis to multivariate representation, recurrent population dynamics, model-based imaging, and task-trained artificial networks. These advances supply different but compatible ways of asking how a biological system organizes cognition and behavior. The evidence demonstrates that explicit models can connect behavioral and neural measurements in powerful ways, while the exact character of each correspondence depends on task, observation process, parameterization, and experimental design. The scientific unit of progress is a tested bridge hypothesis with an identifiable scope. Hodgkin & Huxley (1952) Marr & Poggio (1976) Kriegeskorte & Douglas (2018)

The strongest conclusions are those that withstand multiple forms of constraint. A model may explain behavior, representational geometry, trial-wise neural variation, and population dynamics; an intervention may then test whether its proposed causal process participates in producing action. Each outcome answers a specific question, and disagreement among them can reveal which part of the theory needs revision. The principal methodological priorities are careful behavioral decomposition, identifiable computational variables, faithful neural measurement models, comparisons with substantive alternatives, independent generalization, and selectively designed perturbations. Such priorities make the research cumulative and reduce the risk of treating computational resemblance as a fully established neural mechanism. Krakauer et al. (2017) Diedrichsen & Kriegeskorte (2017) Poldrack (2006)

The historical and contemporary literature supports a disciplined integration of psychology, physiology, mathematics, and computational modeling. Biological experiments constrain implementations, computational models formalize possible processes, and psychological research identifies the behavior that explanation must account for. Artificial networks enrich this interaction through constructive demonstration and controlled comparative experiments, while their relationship to biological cognition remains a question for measurement. Cross-Level Model Mapping therefore names a continuing empirical task: to specify and test the transformations that connect explanatory variables across scales, preserving the distinction between prediction, representation, and causal realization as the research develops. Yamins et al. (2014) Hassabis et al. (2017) Saxe et al. (2021)

References

  • Abbott, L. F., & Dayan, P. (2001). Theoretical neuroscience: Computational and mathematical modeling of neural systems. MIT Press. https://mitpress.mit.edu/9780262041997/theoretical-neuroscience/
  • Churchland, M. M., Cunningham, J. P., Kaufman, M. T., Foster, J. D., Nuyujukian, P., Ryu, S. I., & Shenoy, K. V. (2012). Neural population dynamics during reaching. Nature, 487, 51–56. https://doi.org/10.1038/nature11129
  • Diedrichsen, J., & Kriegeskorte, N. (2017). Representational models: A common framework for understanding encoding, pattern-component, and representational-similarity analysis. PLOS Computational Biology, 13(4), e1005508. https://doi.org/10.1371/journal.pcbi.1005508
  • Ding, Z., Fahey, P. G., Papadopoulos, S., Wang, E. Y., Celii, B., Papadopoulos, C., Chang, A., Kunin, A. B., Tran, D., Fu, J., Ding, Z., Patel, S., Ntanavara, L., Froebe, R., Ponder, K., Muhammad, T., Bae, J. A., Bodor, A. L., Brittain, D., ... Tolias, A. S. (2025). Functional connectomics reveals general wiring rule in mouse visual cortex. Nature, 640, 459–469. https://doi.org/10.1038/s41586-025-08840-3
  • Dorkenwald, S., Matsliah, A., Sterling, A. R., Schlegel, P., Yu, S.-C., McKellar, C. E., Lin, A., Costa, M., Eichler, K., Yin, Y., Silversmith, W., Schneider-Mizell, C., Jordan, C. S., Brittain, D., Halageri, A., Kuehner, K., Ogedengbe, O., Morey, R., Gager, J., ... The FlyWire Consortium. (2024). Neuronal wiring diagram of an adult brain. Nature, 634, 124–138. https://doi.org/10.1038/s41586-024-07558-y
  • Friston, K. (2010). The free-energy principle: A unified brain theory? Nature Reviews Neuroscience, 11, 127–138. https://doi.org/10.1038/nrn2787
  • Glaser, J. I., Benjamin, A. S., Farhoodi, R., & Kording, K. P. (2019). The roles of supervised machine learning in systems neuroscience. Progress in Neurobiology, 175, 126–137. https://doi.org/10.1016/j.pneurobio.2019.01.008
  • Hassabis, D., Kumaran, D., Summerfield, C., & Botvinick, M. (2017). Neuroscience-inspired artificial intelligence. Neuron, 95(2), 245–258. https://doi.org/10.1016/j.neuron.2017.06.011
  • Haxby, J. V., Gobbini, M. I., Furey, M. L., Ishai, A., Schouten, J. L., & Pietrini, P. (2001). Distributed and overlapping representations of faces and objects in ventral temporal cortex. Science, 293(5539), 2425–2430. https://doi.org/10.1126/science.1063736
  • Hodgkin, A. L., & Huxley, A. F. (1952). A quantitative description of membrane current and its application to conduction and excitation in nerve. The Journal of Physiology, 117(4), 500–544. https://doi.org/10.1113/jphysiol.1952.sp004764
  • Krakauer, J. W., Ghazanfar, A. A., Gomez-Marin, A., MacIver, M. A., & Poeppel, D. (2017). Neuroscience needs behavior: Correcting a reductionist bias. Neuron, 93(3), 480–490. https://doi.org/10.1016/j.neuron.2016.12.041
  • 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
  • Kriegeskorte, N., & Douglas, P. K. (2018). Cognitive computational neuroscience. Nature Neuroscience, 21, 1148–1160. https://doi.org/10.1038/s41593-018-0210-5
  • Logothetis, N. K., Pauls, J., Augath, M., Trinath, T., & Oeltermann, A. (2001). Neurophysiological investigation of the basis of the fMRI signal. Nature, 412, 150–157. https://doi.org/10.1038/35084005
  • Mante, V., Sussillo, D., Shenoy, K. V., & Newsome, W. T. (2013). Context-dependent computation by recurrent dynamics in prefrontal cortex. Nature, 503, 78–84. https://doi.org/10.1038/nature12742
  • Marr, D., & Poggio, T. (1976). From understanding computation to understanding neural circuitry (AIM-357). MIT Artificial Intelligence Laboratory. https://dspace.mit.edu/entities/publication/c6722773-6065-4c2c-998c-7190be63f19a
  • MICrONS Consortium. (2025). Functional connectomics spanning multiple areas of mouse visual cortex. Nature, 640, 435–447. https://doi.org/10.1038/s41586-025-08790-w
  • Nili, H., Wingfield, C., Walther, A., Su, L., Marslen-Wilson, W., & Kriegeskorte, N. (2014). A toolbox for representational similarity analysis. PLOS Computational Biology, 10(4), e1003553. https://doi.org/10.1371/journal.pcbi.1003553
  • O'Doherty, J. P., Hampton, A., & Kim, H. (2007). Model-based fMRI and its application to reward learning and decision making. Annals of the New York Academy of Sciences, 1104, 35–53. https://doi.org/10.1196/annals.1390.022
  • Poldrack, R. A. (2006). Can cognitive processes be inferred from neuroimaging data? Trends in Cognitive Sciences, 10(2), 59–63. https://doi.org/10.1016/j.tics.2005.12.004
  • Rao, R. P. N., & Ballard, D. H. (1999). Predictive coding in the visual cortex: A functional interpretation of some extra-classical receptive-field effects. Nature Neuroscience, 2, 79–87. https://doi.org/10.1038/4580
  • 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
  • Schultz, W., Dayan, P., & Montague, P. R. (1997). A neural substrate of prediction and reward. Science, 275(5306), 1593–1599. https://doi.org/10.1126/science.275.5306.1593
  • Sussillo, D., Churchland, M. M., Kaufman, M. T., & Shenoy, K. V. (2015). A neural network that finds a naturalistic solution for the production of muscle activity. Nature Neuroscience, 18, 1025–1033. https://doi.org/10.1038/nn.4042
  • 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., Joglekar, M. R., Song, H. F., Newsome, W. T., & Wang, X.-J. (2019). Task representations in neural networks trained to perform many cognitive tasks. Nature Neuroscience, 22, 297–306. https://doi.org/10.1038/s41593-018-0310-2