Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Functional MRI and Multivariate Neural Decoding

Abstract. Functional MRI and multivariate neural decoding investigate how magnetic resonance signals associated with neurovascular responses can be analyzed to predict experimentally specified perceptual, cognitive and behavioral variables. This fundamental research monograph critically integrates the origins and physics of blood-oxygen-level-dependent (BOLD) imaging, acquisition and task design, general linear modeling, multivoxel pattern analysis, encoding and decoding, representational similarity, searchlight mapping, and cross-validated statistical inference. Primary experimental studies on visual representations, natural-image identification, semantic reconstruction, cross-modal language representation and autobiographical imagery are evaluated alongside methodological investigations of circular analysis, participant-level generalization, analytical flexibility, data sharing and reproducibility. The evidence supports conditional recovery of information from distributed brain measurements while demanding precise distinctions among task prediction, psychological construct measurement, causal explanation and generative-model prior. It identifies unresolved questions about vascular specificity, population transport, independent replication, prospective clinical benefit and informed governance of mental information. The study is a critical narrative scientific synthesis of the published literature; it reports no original participant data, registered systematic-review protocol, newly calculated meta-analysis or individualized clinical advice.

Keywords: functional magnetic resonance imaging; fMRI; BOLD signal; neurovascular coupling; multivoxel pattern analysis; MVPA; neural decoding; neural encoding; representational similarity analysis; searchlight analysis; cross-validation; nested cross-validation; reverse inference; brain reading; semantic reconstruction; naturalistic cognition; generative brain decoding; reproducible neuroimaging; brain-computer interfaces; mental privacy

Introduction: The Scientific Object of Functional MRI and Neural Decoding

Functional magnetic resonance imaging (fMRI) measures changes in magnetic resonance signals associated with physiology during experimentally structured or naturally occurring brain activity. Multivariate neural decoding uses patterns sampled across measurement channels to predict independently specified stimuli, behaviors, tasks, or representational variables. Together these methods have transformed cognitive neuroscience by allowing investigators to ask which distinctions are recoverable from distributed human brain responses, where those distinctions are expressed, and how reliably they persist across observations. The scientific object of this monograph is the measurement-and-inference chain connecting magnetic resonance signal generation, experimental design, statistical estimation, representational analysis, and psychological interpretation. It concerns the evidence conveyed by measured signals and the assumptions under which that evidence can support claims about cognition. Ogawa et al. (1990) Norman et al. (2006) Hebart & Baker (2018)

The central distinction is between a measurable physiological contrast, an informative statistical pattern, and a scientifically warranted explanation. A task-related BOLD difference can support an association between a condition and a hemodynamic response; successful cross-validated classification can establish that a predefined target is statistically discriminable from held-out measurements. Neither result alone identifies a particular mental mechanism, causal neural code, or literal subjective thought. The variables available to a decoder may include stimulus properties, decisions, motor preparation, attention, movement, breathing, and apparatus-related features. Valid inference therefore depends on explicit construct definition, experimental controls, and an accurate characterization of how information enters a learning algorithm. Poldrack (2006) Hebart & Baker (2018) Davis et al. (2014)

This work is a critical narrative synthesis of original BOLD-physics studies, foundational fMRI and pattern-decoding experiments, methodological research, independent critiques, reproducibility investigations, and recent generative decoding demonstrations. It evaluates the distinction between subject-specific and population-level prediction; stimulus decoding and reconstruction; localization and distributed information; explanation and prediction; and technical performance and broader ethical consequences. No participant data are collected, no systematic-review protocol or pooled meta-analysis is claimed, and the empirical findings reported here remain attributable to the original investigators. A decisive criterion governs the discussion: prediction is scientifically valuable when its validation design and target support the particular interpretation advanced. Haynes & Rees (2006) Nichols et al. (2017) Marek et al. (2022)

Definitions, Research Questions, and Levels of Explanation

Blood-oxygen-level-dependent (BOLD) contrast denotes an MRI signal dependence on the balance of oxygenated and deoxygenated blood, vascular properties, and the magnetic properties of deoxyhemoglobin. A voxel is a spatial sampling element containing signal contributions from many microscopic components, including vessels and tissue compartments. A condition contrast is an estimated difference between experimental conditions after accounting for modeled nuisance variables. A multivoxel pattern is a vector of responses indexed by measurement locations. A decoder is a fitted mapping from such a vector to a target label or quantity, assessed on data unavailable to the fitting procedure. These definitions clarify why a voxel has no one-to-one identity with a neuron, and why a pattern is a representation in an analysis space rather than direct access to internal experience. Ogawa et al. (1990) Logothetis et al. (2001) Naselaris et al. (2011)

Neural encoding models estimate how a specified stimulus or cognitive feature predicts measured responses; decoding models invert the predictive direction by estimating the target from observed brain data. Both can be useful even when the causal mechanisms are incompletely specified. Linear classifiers seek separating functions, whereas encoding models can quantify a stimulus feature basis and test how predicted responses change with novel inputs. Representational similarity analysis evaluates relationships among condition patterns rather than relying on one label decision. A searchlight repeats an information-sensitive statistic across local neighborhoods to create a map of available discriminative information. The choice of method determines what scientific question a positive result can answer. Kriegeskorte et al. (2006) Kriegeskorte et al. (2008) Naselaris et al. (2011)

The principal questions are where predictive information can be measured, which aspect of cognition the prediction concerns, how estimates generalize to new runs, sessions, subjects, scanners and tasks, and how alternative physiological or behavioral explanations can be excluded. A second family of questions concerns comparison across theories: which representational geometry, population mechanism, or computational feature basis best predicts held-out data? A third concerns ethical and practical translation: whether prediction is reliable enough to serve an application, how consent and mental privacy are secured, and which apparent capacities arise from intensive training or generative priors rather than signal-alone reconstruction. Each question requires a different design and error model. Haxby et al. (2011) Tang et al. (2023) Poldrack et al. (2017)

Origins of Functional Magnetic Resonance Brain Mapping

The origins of functional MRI combine magnetic-resonance physics, physiological contrast, human sensory mapping, and statistical image analysis. Ogawa and colleagues characterized the oxygenation dependence of magnetic resonance contrast in 1990, identifying deoxyhemoglobin as an intrinsic source of susceptibility contrast. Belliveau and colleagues published functional mapping of human visual cortex by MRI in 1991 using an injected contrast-agent approach, illustrating the importance of blood-flow changes for task mapping. Kwong and colleagues demonstrated dynamic imaging of human responses to sensory stimulation in 1992 with techniques including endogenous contrast. These milestones describe distinct experimental routes and should not be compressed into a single claim that one paper invented every component of modern BOLD fMRI. Ogawa et al. (1990) Belliveau et al. (1991) Kwong et al. (1992)

Early sensory experiments established that time-varying signals could be related to controlled visual and other stimulation and localized through repeated acquisition. Boynton and colleagues subsequently tested a linear-systems characterization of the functional response in primary visual cortex, making explicit the conditions under which neural events can be convolved with a canonical hemodynamic response to predict measured time courses. The rise of event-related experimental designs made task timing, hemodynamic latency, temporal correlation, and statistical power central engineering questions. The apparent visual simplicity of colored activation maps conceals a complex sequence from raw phase and frequency encoding to physiological modeling and multiple-comparison control. Boynton et al. (1996) Woolrich et al. (2001)

A second historical development concerned the meaning of a regional activation. Classical contrast maps test whether a modeled condition is associated with responses at particular locations; they do not exhaust the possible information carried jointly by responses spread across those locations. Haxby and colleagues showed in 2001 that response patterns across ventral temporal cortex distinguished visual categories even after maximally category-responsive regions were excluded. Cox and Savoy later developed explicit classification of distributed visual-cortex patterns, and Kamitani and Tong demonstrated fine-grained decoding of visual orientation and the subjective perception of ambiguous stimuli. This sequence shifted the question from which isolated voxel responds most to which distinctions can be decoded from a population measurement. Haxby et al. (2001) Cox & Savoy (2003) Kamitani & Tong (2005)

BOLD Contrast, Neurovascular Coupling, and Signal Formation

The physical basis of BOLD imaging is the magnetic susceptibility of deoxygenated hemoglobin and its influence on local transverse relaxation and image intensity. Neural activity influences local energy demand and vascular regulation; the resulting blood-flow response can overcompensate for oxygen consumption, changing venous deoxyhemoglobin and measured contrast. Because the MRI signal integrates multiple vascular and tissue compartments, the relationship between local neural events and an observed amplitude depends on vascular reactivity, baseline physiology, sequence parameters, field strength, and spatial location. BOLD images therefore measure a mediated physiological consequence of neural processing. Quantitative interpretation requires a model of that mediation. Ogawa et al. (1990) Buxton et al. (1998) Logothetis et al. (2001)

Buxton, Wong, and Frank introduced a dynamic model of flow and oxygenation that became known as the balloon model, clarifying how blood volume and deoxyhemoglobin changes shape the hemodynamic time course. Boynton and colleagues' linear-systems experiments offered an operational approximation for responses to temporal stimulation in human visual cortex. These models have different objectives: one characterizes vascular dynamics, and the other tests predictability under simplified convolution assumptions. Neither licenses the assumption that a change in BOLD signal is a direct measurement of spike count or a unique signature of excitation. Spatially varying delays, nonlinearities, habituation, and vascular physiology complicate interpretation, particularly for comparisons across regions or populations. Buxton et al. (1998) Boynton et al. (1996)

Logothetis and colleagues conducted simultaneous electrophysiological and fMRI measurements in monkey visual cortex and reported a substantial relationship between BOLD responses and local field potentials, motivating a closer account of synaptic and population activity than simplistic spike-based metaphors allow. The finding is important precisely because physiological correspondence is conditional on recording circumstances and cannot settle the relationship in every human task or brain structure. A successful decoder can exploit stable hemodynamic covariation without isolating the neural generator of that covariation. For research involving disease, age, pharmacological exposure, arousal, or vascular differences, an explicit assessment of neurovascular confounding becomes an integral part of construct validity. Logothetis et al. (2001) Davis et al. (2014) Nichols et al. (2017)

Acquisition, Spatial Sampling, and Temporal Limitations

Echo-planar imaging samples repeated brain volumes at an acquisition interval chosen in relation to field of view, spatial resolution, signal-to-noise, and practical scanning constraints. Spatial smoothing, interpolation, head-motion correction, spatial normalization, and magnetic susceptibility distortion correction all alter the relation between raw signal and final analysis features. A nominally fine voxel grid does not confer neuronal spatial specificity: physiological drainage, point-spread effects, and preprocessing create shared variation across locations. Conversely, temporally sampled images provide a slow hemodynamic view of neural events that may evolve on much shorter time scales. Scientific claims about individual stimulus timing or direction-selective circuits need to account for those measurement scales. Boynton et al. (1996) Esteban et al. (2019) Logothetis et al. (2001)

Acquisition choices are integral to the downstream decoding estimator. High-dimensional patterns allow many weakly informative locations to contribute to a separating decision; that flexibility also magnifies the opportunity for nuisance information to enter classification. Motion correlated with task labels, differences in temporal order, scanner drift, and physiological fluctuations may be sufficient to support prediction. Sensory and motor confounds must be addressed at the experimental design stage because an algorithm trained on a confounded dataset can produce stable, replicable accuracy for the wrong explanatory reason. The evidential unit is the whole acquisition–task–model system and not a decoder evaluated separately from its experimental context. Cox & Savoy (2003) Davis et al. (2014) Nichols et al. (2017)

Reusable workflows make image acquisition and transformation more transparent. The Brain Imaging Data Structure specifies conventions for organizing neuroimaging datasets and their essential metadata, supporting exchange and downstream reproducibility. fMRIPrep provides a standardized preprocessing pipeline while leaving scientific design decisions and statistical analysis to the investigator. Such tools reduce arbitrary variation in foundational data handling, but they do not establish that motion has been adequately controlled, that a particular spatial mask is appropriate, or that a classification result will transport across datasets. Quality-control reports, retained acquisition metadata, explicit censoring or nuisance models, and task timing files remain necessary scientific evidence. Gorgolewski et al. (2016) Esteban et al. (2019) Poldrack et al. (2017)

From Task Design to Statistical Contrasts

Task-based fMRI starts with explicit psychological manipulations, controls, and measurement outcomes. A contrast between two tasks estimates response differences conditional on an experimental model; interpreting that difference as a component of cognition requires showing which demands vary and which remain controlled. Block designs often provide favorable detection power but increase the opportunity for slowly changing vigilance or movement to covary with labels. Event-related designs separate trial types more flexibly but demand attention to timing, overlap of hemodynamic responses, and the reliability of single-trial estimates. The strongest experimental design depends on whether the target is a condition-average effect, a trial-level behavioral variable, or a generalized response to a novel stimulus. Boynton et al. (1996) Woolrich et al. (2001) Norman et al. (2006)

The general linear model expresses a measured time series as a combination of hypothesized task responses and nuisance terms plus error. Coefficients can provide interpretable subject-level condition estimates, which may be used as decoder features in later analyses. Temporal autocorrelation makes residuals dependent over time, so treating successive volumes as independent observations exaggerates available information. Woolrich and colleagues examined methods for handling autocorrelation in univariate functional imaging, but the statistical concern extends to multivariate feature extraction and evaluation. If a train–test split places adjacent, autocorrelated samples on opposite sides of the split, even a technically held-out volume may carry information leaked from its training neighbors. Woolrich et al. (2001) Varoquaux (2018)

Population inference requires additional care because repeated volumes within a person and repeated runs within a session are not independent participants. Fixed-effects questions about a particular measured sample, random-effects claims about an underlying participant population, and predictive claims about new people invoke distinct sampling units. A group-level significant contrast does not imply that a classifier works reliably for every participant; a highly accurate within-person decoder does not imply that it generalizes to a new individual. A rigorous report defines both the measurement unit and the intended population of inference in advance, then arranges experimental replication at that level. Allefeld et al. (2016) Marek et al. (2022) Nichols et al. (2017)

Distributed Representation and the Emergence of Multivoxel Pattern Analysis

Multivoxel pattern analysis (MVPA) treats a collection of fMRI measurements as a jointly informative signal. Its central premise is that a task or stimulus distinction may be present in the relative configuration of many weak responses even when no single voxel has a large mean contrast. Haxby and colleagues' 2001 object-category study provided influential evidence for distributed, overlapping ventral temporal representations. Cox and Savoy's 2003 classification work and the synthesis by Norman and colleagues in 2006 helped consolidate pattern-based prediction as an experimental method. These results expanded what could be studied with existing imaging hardware while also introducing the inferential challenge of interpreting weights and successful predictions. Haxby et al. (2001) Cox & Savoy (2003) Norman et al. (2006)

A distributed pattern can contain category information without exhibiting a neat anatomical boundary. Information may reside in mean differences, the geometry of condition responses, interaction terms, or correlations arising from common physiological inputs. Classification performance can therefore remain positive when the feature set has no simple functional-module interpretation. Davis and colleagues analyzed the relationships between multivoxel and univariate analyses, showing why divergent results can reflect distinct variance components rather than an automatic discovery of a more fundamental code. Any claim that a decoder reveals a new neural representation should identify which features of the signal yield the result and compare appropriate univariate and nuisance-variable controls. Davis et al. (2014) Hebart & Baker (2018)

The word 'reading' can be a useful shorthand for prediction under a specified experimental distribution, but it is scientifically misleading when it suggests unrestricted access to unprompted thought. In classical MVPA, a participant performs an experiment, the investigator chooses target variables, and the classifier learns correlations from labeled training data. Information is assessed by whether those learned associations predict new observations under a designed test. Improvements in classification methodology do not eliminate these prerequisites. The strongest interpretations rely on a transparent task, an independently defined target, a well-controlled validation design, and a quantitative description of error rather than a metaphor of machine telepathy. Haynes & Rees (2006) Poldrack (2006) Tang et al. (2023)

Classification Algorithms, Feature Spaces, and Decision Boundaries

Linear classifiers, including regularized logistic regression and linear support vector machines, combine voxel responses through fitted weights to separate conditions or predict a label. Regularization controls model complexity and is important when the number of voxels is large relative to the number of independent observations. More flexible nonlinear estimators can represent additional interactions, but their added capacity increases demands on data size and out-of-sample evaluation. The scientific issue is not which algorithm produces the most impressive training accuracy; it is which class of models generalizes and which patterns of error distinguish hypotheses about neural coding. Comparisons should hold preprocessing, splits, and model-selection budgets constant. Misaki et al. (2010) Hebart & Baker (2018)

Normalization is a substantive analytic choice. Centering a voxel within training observations, scaling variance, detrending, removing run means, and estimating nuisance parameters can change the geometry a classifier sees. Misaki and colleagues compared classifiers and response-normalization approaches for pattern-information fMRI, illustrating that these decisions affect sensitivity. A principled pipeline estimates any transform from the training partition and applies it unchanged to test observations. Estimating means, feature selection, or scaling parameters from all observations before cross-validation can leak information about test distributions into training and undermine the intended independence. This is particularly serious when a small dataset invites complex hyperparameter searches. Misaki et al. (2010) Kriegeskorte et al. (2009) Varoquaux (2018)

Decoding accuracy should be reported alongside the target prevalence, class balance, confusion patterns, confidence intervals, and suitable baseline models. An apparent improvement over a simple majority-class predictor may be irrelevant if performance is unstable across scans or if labels correlate with movement or temporal order. For continuous targets, investigators may use held-out correlation, predictive R-squared, error magnitude, or likelihood; each answers a different question about prediction and calibration. Model weights are conditional on correlated features and regularization, and large weights do not necessarily indicate essential causal or anatomically specific brain regions. Perturbation, ablation, feature-stability analysis, and comparison of alternative representational accounts are needed for stronger explanatory claims. Davis et al. (2014) Hebart & Baker (2018) Nichols et al. (2017)

Cross-Validation, Independent Tests, and Leakage Control

A decoder's central evidential claim is its behavior on genuinely unseen data. Cross-validation partitions observations into training and test sets to estimate this behavior, but its validity depends on whether the partition matches the independence structure and intended deployment. Randomly splitting temporally adjacent MRI volumes can transmit shared hemodynamic dynamics or run-specific noise across the split. Leave-one-run-out validation is often more defensible for within-person task decoding, whereas leave-one-session-out tests scanner-day stability and leave-one-participant-out protocols address a different population. None is universally sufficient: a model intended for a different scanner, age cohort, or language task requires validation that withholds the corresponding source of variation. Varoquaux (2018) Woolrich et al. (2001) Nichols et al. (2017)

The entire model-selection procedure belongs inside training data. If a researcher compares hundreds of feature sets or classifier parameters and reports the best on the same held-out set, that set has effectively become part of model development. Nested cross-validation separates inner tuning from outer performance estimation, while a genuinely external test sample provides a clearer audit of fixed modeling choices. Selection of regions of interest by looking at the target contrast in the full dataset creates a related circularity. Kriegeskorte and colleagues demonstrated that 'double dipping' can manufacture apparently meaningful findings even in data lacking the target effect. Independence must be established at the level of the scientific contrast as well as in code. Kriegeskorte et al. (2009) Varoquaux (2018)

Varoquaux's investigation of small-sample cross-validation emphasized that uncertainty in estimated generalization can remain large even when standard fold-wise variability looks modest. Reported folds share training observations; treating fold-level scores as independent replications can produce unjustifiably narrow uncertainty intervals. Repeated partitioning, bootstrap procedures appropriate to the dependence structure, and external test cohorts can help characterize instability, although none repairs an invalid sampling design. Researchers should predefine the principal performance metric, disclose all attempted models, and distinguish exploratory selection from confirmatory testing. For population inference about the presence of information, the statistical null and analysis units require separate attention beyond simple above-chance classifications. Varoquaux (2018) Allefeld et al. (2016) Poldrack et al. (2017)

Searchlight Mapping, Regional Localization, and Informational Sensitivity

Information-based searchlight mapping moves a spatial neighborhood through an image and estimates an information statistic for each location. Kriegeskorte, Goebel, and Bandettini introduced a prominent formulation in 2006. The resulting map answers where neighborhoods of voxels contain information sufficient for the chosen discrimination, conditional on the classifier and validation procedure. It differs from a univariate activation map because discriminability may arise from joint patterns of responses. A searchlight center is therefore a convenient location for reporting the neighborhood statistic; it is not a demonstration that the central voxel alone carries the effect or that the underlying code is anatomically confined to the neighborhood. Kriegeskorte et al. (2006)

Neighborhood radius, voxel geometry, spatial smoothing, cortical folding, signal-to-noise, and the density of informative features all affect sensitivity. Two regions might differ in searchlight accuracy because one has more reliable measurement, not because its neuronal representation is richer. Searchlight maps and region-of-interest analyses can complement one another, but they make different trade-offs between spatial exploration and hypothesis-driven sensitivity. An appropriately specified independent ROI avoids the inflated selectivity that arises when the same data determine the chosen region and test its response. Separate localizer runs, anatomical definitions, or wholly nested selection are options whose validity depends on the target claim and the independence of relevant noise. Kriegeskorte et al. (2009) Misaki et al. (2010)

Mass-univariate inference and searchlight-based maps share a multiple-testing burden, even though their test statistics differ. Spatial permutations, exchangeability-aware shuffles, and prespecified correction procedures can assess whether an information map exceeds what would occur under a valid null. Temporal dependence, repeated measurements, and experimental blocks restrict which labels can legitimately be permuted. Spatial clusters or significant voxels should be interpreted in relation to the exact inferential statistic rather than described as direct maps of isolated neural concepts. Convergent findings across independent tasks, independent samples, and complementary imaging modalities provide stronger evidence of generalizable functional organization than a single, optimized spatial map. Kriegeskorte et al. (2006) Nichols et al. (2017) Poldrack et al. (2017)

Visual Feature Decoding and the Limits of Spatial Granularity

Visual decoding studies are central because investigators can tightly manipulate stimulus features while measuring behavioral reports. Kamitani and Tong demonstrated that information about perceived edge orientation and perceptual interpretation could be recovered from human visual-cortex fMRI patterns. Their later investigation of seen and attended motion directions extended the logic to feature-selective attention. These experiments show that a pattern carries information about task-relevant visual distinctions under the tested conditions. They do not require the imaging voxel itself to be orientation-selective at the single-neuron scale; spatial biases and aggregation can make a finer underlying arrangement accessible to a coarse measurement. Kamitani & Tong (2005) Kamitani & Tong (2006)

The methodological lesson concerns the relationship between representational granularity and spatial sampling. A classifier may discriminate stimulus categories by exploiting coarse-scale biases that emerge from many underlying neurons, retinotopic differences, vascular architecture, or stable measurement anisotropies. The presence of decodable information therefore underdetermines which microscopic population code produced it. Comparing stimuli that preserve low-level image properties, altering retinotopic position, and evaluating transfer across sessions can test alternative accounts. Conversely, failure to classify a particular orientation at a given resolution does not demonstrate the absence of orientation-selective neural populations. Both successes and failures are conditioned by acquisition sensitivity and model assumptions. Kamitani & Tong (2005) Logothetis et al. (2001) Hebart & Baker (2018)

An important distinction separates task performance from subjective report. When ambiguous stimuli yield changes associated with reported perception, a decoder may supply evidence about percept-dependent patterns. Still, reports introduce their own motor, decision, and expectation signals, so interpretation becomes stronger with carefully controlled response mappings and converging experimental contrasts. The scientific target is a relationship between the independently operationalized perceptual state and the observed multivariate signal. This is more precise than treating decoded labels as unrestricted observation of consciousness. A robust experimental program varies physical input and behavioral report to determine which components of the signal track stimulus encoding, attention, choice, or perceptual experience. Kamitani & Tong (2005) Poldrack (2006)

Encoding Models, Representational Similarity, and Hyperalignment

An encoding model specifies a mapping from stimulus features to predicted brain responses. Kay and colleagues developed quantitative receptive-field models to identify natural images from human visual-cortex activity, illustrating how a feature-based encoding model can support a subsequent identification procedure. Naselaris and colleagues clarified the complementary uses of encoding and decoding: the former articulates a candidate stimulus-to-response transformation, while the latter evaluates what information about a target is accessible from responses. Encoding models permit more direct tests between representational hypotheses because alternative feature spaces can be used to generate predictions for novel stimuli. Their explanatory power remains constrained by how features were chosen and whether competing models were compared on identical held-out observations. Kay et al. (2008) Naselaris et al. (2011)

Representational similarity analysis (RSA) studies the geometry of pairwise differences among condition patterns. Kriegeskorte, Mur, and Bandettini described an approach in which brain representational dissimilarity matrices can be compared with behavioral measurements or computational models without assuming a voxel-by-voxel correspondence to model units. The result can reveal whether similar experimental objects evoke similar measured patterns and whether a proposed model captures that relational structure. The validity of the comparison depends on noise estimation, choice of dissimilarity measure, stimulus sampling, and generalization to new conditions. Similarity of geometries is evidence about representational organization, and causal implementation remains a separate question. Kriegeskorte et al. (2008) Hebart & Baker (2018)

Representations across participants may be functionally similar even when voxel-by-voxel anatomical correspondence is weak. Haxby and colleagues introduced a common high-dimensional representational space for ventral temporal cortex and demonstrated the utility of aligning fine-grained patterns across people. This hyperalignment approach helped establish that functional correspondence can require transformations beyond standard anatomical registration. Its success depends on common stimulus structure, reliable responses, and the generalization of the alignment beyond the material used to estimate it. Training alignment on the same conditions that later define the tested hypothesis can contaminate evaluation unless the pipeline withholds the appropriate samples. Cross-participant decoding thus combines a neuroscientific claim about shared structure with a statistical claim about out-of-sample transport. Haxby et al. (2011) Varoquaux (2018) Nichols et al. (2017)

Natural Images, Feature-Based Identification, and Generative Reconstruction

The progression from simple orientation and category labels to complex natural stimuli required richer predictive models and better defined stimulus spaces. Kay and colleagues showed how quantitative visual encoding models could be used to identify previously unseen natural images from a candidate set using measured visual-cortex responses. This constituted a landmark in model-based identification, but the candidate set, experimental calibration, and visual feature representation were crucial components of the procedure. Identification among alternatives is a distinct statistical problem from generating a uniquely correct perceptual reconstruction. The difference matters because a model can select the best candidate successfully without preserving every detail of a remembered or perceived scene. Kay et al. (2008) Naselaris et al. (2011)

Generative image reconstruction introduces a second model layer: the measured pattern conditions a high-capacity image generator already trained on extensive visual data. Takagi and Nishimoto presented a latent-diffusion approach to reconstructing images from fMRI activity, demonstrating the feasibility of combining brain-derived features with a learned image prior. Reconstructions can be visually compelling because the generator supplies plausible structure under constraints supplied by the brain data. Consequently, photorealism alone cannot establish which details were specified by the measured signal. Evaluation should distinguish semantic agreement, low-level feature preservation, candidate ranking, trial-level reproducibility, and the extent to which results change when prior knowledge or model components are altered. Takagi & Nishimoto (2023) Naselaris et al. (2011)

A rigorous reconstruction study separates the training of a generic generative model, the estimation of participant-specific brain mappings, and the evaluation of new test stimuli. The relevant baseline might include predictions based on image category or text description alone, a shuffled brain pattern, a participant-independent mapping, and reconstruction under alternative priors. Novel stimulus sets should contain carefully controlled semantic neighbors so that coarse labels do not masquerade as detailed recovery. Repeated scanning of the same conditions can quantify reproducibility, while blind human raters and clearly defined automated metrics can assess different dimensions of similarity. The scientific conclusion is strongest when successful generalization survives these alternatives and when qualitative examples are accompanied by the full distribution of results. Kay et al. (2008) Takagi & Nishimoto (2023) Poldrack et al. (2017)

Language, Semantic Decoding, and Naturalistic Cognition

Naturalistic decoding of speech, narrative, and semantic content poses a different challenge from classifying tightly controlled stimuli. Language unfolds over time, combines multiple levels of meaning, and draws on wide networks whose BOLD responses are temporally blurred. Tang and colleagues reported a noninvasive semantic decoder that reconstructed intelligible sequences conveying aspects of the meaning of heard stories, imagined speech, and silent videos. The reported approach combined extensive individualized fMRI training with language-model-based candidate generation and evaluation. Its findings establish a significant proof of concept for semantic reconstruction under specified acquisition and cooperation conditions. Tang et al. (2023)

The phrase 'reconstructs language' must be understood at the level actually assessed. A semantic output can preserve gist while replacing exact wording, and its fluency partly reflects constraints supplied by a pretrained language model. Training and evaluation were also dependent on individual cooperation, which the investigators examined explicitly in resistance experiments. This differs from arbitrary extraction of private speech from a stranger encountered in everyday life. Scientifically, the result raises questions about which semantic features are observable in the cortical data, how much linguistic structure derives from the model prior, and how decoding accuracy changes with training duration, subject identity, genre, topic, and communicative intent. Tang et al. (2023) Haynes & Rees (2006)

Naturalistic protocols can reveal aspects of cognition that highly artificial trials neglect, but they complicate causal interpretation. Spoken narratives simultaneously vary semantics, prosody, imagery, attention, autobiographical associations, and memory. Computational feature models can estimate partial contributions if the stimulus ensemble supplies enough independent variation. Carefully designed counterfactual materials, repeated listening, cross-story transfer, semantic controls, and comparison across encoding feature spaces improve explanatory specificity. Naturalistic research therefore requires an explicit distinction between decoding a useful statistical correlate of meaning and identifying a uniquely necessary semantic code. Its promise lies in expanding the range of observable psychological phenomena while preserving the experimental discipline that made classical task decoding interpretable. Naselaris et al. (2011) Tang et al. (2023) Poldrack (2006)

Psychological Reverse Inference and Construct Validity

Reverse inference occurs when activation of a brain region or the presence of a measured pattern is used to infer that a particular cognitive process was engaged. Poldrack's analysis showed that this reasoning depends on how selectively the signal is associated with the process across alternative tasks. A region participating in many cognitive operations offers limited diagnostic evidence about any one operation. Multivariate classifiers can provide more specific prediction than a single-region label, yet they remain vulnerable to the same construct problem if training labels fail to isolate the psychological state of interest. The essential distinction is between a predictive mapping to an experimentally assigned category and an empirically validated measure of a mental construct. Poldrack (2006) Haynes & Rees (2006)

A decoder's target may be a stimulus category, a keypress, an explicit rating, or an investigator-imposed label. These variables need not coincide with perception, intention, memory, or conscious belief. When labels are behaviorally derived, measurement error and task strategy become part of the prediction. When labels are stimulus-derived, subjects may differ in their subjective interpretation of the same item. A study claiming to decode remembered content must therefore establish how actual retrieval or reconstruction was measured and compare it with sensory, attentional, and motor explanations. Construct validation often requires multiple tasks, convergent measures, manipulation checks, and tests showing that the decoder transfers across superficial changes that preserve the proposed cognitive variable. Hebart & Baker (2018) Poldrack (2006) Davis et al. (2014)

Large-scale automated neuroimaging synthesis can provide probabilistic information about the relationships between regions and cognitive task terms. Yarkoni and colleagues established a framework using text mining and large-scale mapping, offering a way to assess broader patterns beyond a single experiment. Such synthesis depends on the vocabulary, task contrasts, reporting conventions, and selection biases of the underlying literature. A classifier trained on literature-derived labels predicts those labels as operationalized by past studies. Strong psychological inference requires additional evidence that the labels correspond to reliable distinctions between mechanisms and that the signals generalize to the new population and context of use. Statistical specificity is not interchangeable with ontological specificity. Yarkoni et al. (2011) Poldrack (2006) Nichols et al. (2017)

Selectivity, Circular Analysis, and Competing Explanations

One of the deepest methodological hazards arises when the dataset used to find informative voxels is also used to estimate the magnitude or significance of their information. Kriegeskorte and colleagues demonstrated how circular selection can distort both activation and pattern-information analyses. The problem is structural: selecting the largest observed correlations and then treating their values as independently tested produces upwardly biased estimates even when the original measurements contain only noise. A positive result therefore requires clarity about which observations were used for hypothesis generation, feature definition, parameter estimation, and final testing. Merely describing a classifier as 'cross-validated' leaves unanswered whether the complete analysis was cross-validated. Kriegeskorte et al. (2009)

Alternative explanations should be specified before interpreting a high decoding score. Features reflecting differences in arousal, motor preparation, ocular movements, respiration, run-specific equipment fluctuations, or block timing can predict experimental labels. If the scientific hypothesis concerns distributed coding of object identity, a more informative design varies category while balancing low-level sensory features, response requirements, and temporal order. Negative control targets can test known nuisance features; cross-condition generalization can establish abstraction from superficial properties; and focused ablations can ask whether signal from a predicted region materially contributes. These controls are stronger when registered prospectively rather than selected after an appealing result emerges. Davis et al. (2014) Poldrack et al. (2017) Hebart & Baker (2018)

An anatomically restricted decoder may appear interpretable, but restriction alone does not establish that all relevant features are neuronal or specific. Vascular drainage, shared task timing, and spatially correlated noise may contribute to local patterns. Conversely, whole-brain models can exploit distributed features without showing that any one brain subsystem is indispensable. Hypothesis testing improves when investigators compare explicit alternatives under a matched data regime, including baseline models, transformations, and spatial masks, and use independent replications. Causal claims about necessity or sufficiency generally require separate perturbation designs; the fMRI decoder itself typically demonstrates information accessibility rather than causal implementation. Logothetis et al. (2001) Kriegeskorte et al. (2006) Poldrack (2006)

Interindividual Generalization, Population Inference, and Statistical Power

Within-person decoding can leverage many trials and stable response mappings in an extensively scanned individual, while between-person prediction must contend with anatomical, physiological, cognitive, and scanner differences. The number of trials is not a substitute for the number of independent participants when a model is intended to generalize to a population. High-dimensional response patterns may be reproducible within a person yet require functional alignment to transfer across individuals. Haxby and colleagues' common representational-space work addressed one route to cross-person correspondence; its results motivate stringent tests using participants and stimuli unavailable for alignment and model selection. Haxby et al. (2011) Varoquaux (2018)

Population-level inference from classification accuracies has special mathematical features. Allefeld, Görgen, and Haynes argued that a standard second-level t-test against chance does not provide the same general population interpretation for bounded information-like measures as for activation differences. Their prevalence-inference framework articulates alternative claims about the fraction of the population showing information. The issue demonstrates why a statistically significant mean accuracy cannot automatically be paraphrased as reliable decodability in most people. Researchers should report individual-level distributions, confidence bounds, sample composition, sampling units, and the particular null hypothesis tested. Reporting only a grand mean may conceal marked heterogeneity or a small highly informative subset. Allefeld et al. (2016)

Marek and colleagues examined reproducibility in brain-wide association studies linking individual differences in imaging and behavioral or mental-health phenotypes. Their large-dataset investigation found that robust estimation of many such population associations required far larger participant samples than those typical of classical task-mapping experiments. The result should be applied with its scope intact: the sample-size requirement for weak between-person associations is not identical to that of repeated controlled measurements within one person. Its relevance to neural decoding lies in showing how target variance, sampling design, expected effect size, model flexibility, and the intended population jointly determine the amount of independent evidence needed for transportable prediction. Marek et al. (2022) Varoquaux (2018) Nichols et al. (2017)

Reliability, Analytical Flexibility, and Reproducible Neuroimaging

Neuroimaging results depend on many decisions: acquisition and quality control, motion correction, registration, nuisance regression, temporal filtering, contrast definition, spatial masking, feature scaling, classifier choice, and statistical thresholding. Botvinik-Nezer and colleagues organized a many-team analysis of the same fMRI dataset and documented substantial variation in outcomes attributable to analytical decisions. The investigation illustrates why a successful single analysis should not be mistaken for inevitability under equally reasonable workflows. Multiverse and specification-curve approaches can expose the sensitivity of a scientific claim to these decisions, provided the explored analytic alternatives are substantively justified and transparently reported. Botvinik-Nezer et al. (2020) Nichols et al. (2017)

Community standards make audit and replication more feasible. Gorgolewski and colleagues' BIDS framework addresses the organization and description of imaging data, while Esteban and colleagues' fMRIPrep pipeline standardizes important preprocessing stages. Nichols and colleagues outlined recommended practices for analysis and sharing, and Poldrack and colleagues surveyed threats to transparent and reproducible neuroimaging. These contributions establish an infrastructure of verifiable inputs, recorded transformations, and inspectable outputs. Reproducibility also depends on archiving the exact analysis code, software versions, feature masks, subject exclusions, model hyperparameters, random seeds, and train–test split assignments, with privacy protections appropriate to human imaging data. Gorgolewski et al. (2016) Esteban et al. (2019) Nichols et al. (2017) Poldrack et al. (2017)

Independent replication has multiple meanings. Computational reproducibility asks whether the same data and code yield the reported result. Measurement reliability asks whether repeated observations of comparable states produce stable estimates. External replication asks whether independent investigators recover the substantive effect under an adequately comparable design. Transportability asks whether a model maintains useful performance under new distributions. Each is essential, and none implies all the others. A code archive may exactly regenerate an overfit classifier; a reliable individual pattern may fail in another scanner; a robust task effect may have little predictive value for a clinical decision. A scientifically mature account presents the relevant uncertainty and level of generalization explicitly. Varoquaux (2018) Marek et al. (2022) Poldrack et al. (2017)

Convergent Measurements and Cross-Modal Representation

Progress in multivariate decoding has made the relationship between fMRI, electrophysiology, and psychological representations a testable object. Wang and colleagues combined fMRI multivoxel analyses with stereoelectroencephalography in research on the visual word form area, reporting evidence that written and spoken language can be distinguished within this region using convergent protocols. The study addresses a functional interpretation that could remain ambiguous from regional response averages alone: a cortical area traditionally associated with written words can also show structured responses to auditory language. Convergence across modalities strengthens the hypothesis that language-related information has a distributed neural basis, while electrode coverage, trial structure, and cross-modal generalization delimit its reach. Wang et al. (2025) Haxby et al. (2001) Hebart & Baker (2018)

Cross-modal studies clarify why the identity of a brain area cannot be established by its customary name. A region labeled for one experimental tradition may encode information that is recruited in multiple tasks, and different analyses may detect distinct kinds of sensitivity. Anatomical overlap between auditory and visual responses is compatible with shared abstract codes, intermingled distinct subpopulations, task-linked feedback, or vascular pooling. Discriminating these accounts requires representational transfer tests, stimulus and semantic controls, and measurements capable of distinguishing temporal and spatial sources. Where direct electrophysiological recordings are available, their temporal resolution can constrain models proposed from fMRI, but intracranial data ordinarily sample selected clinical regions and do not supply whole-brain population coverage. Wang et al. (2025) Logothetis et al. (2001) Naselaris et al. (2011)

The scientific gain from combining techniques arises when their complementary measurement assumptions constrain the same hypothesis. A neural prediction may survive across modalities, tasks, and sessions yet still fail to establish causal necessity. Stimulation and lesion methods add different evidence about what happens when a region or pathway is perturbed; computational encoding models add quantitative tests of which features account for response geometry. A theoretically meaningful synthesis therefore specifies the common target across methods and the inferential contribution of each instrument. It avoids attributing the full interpretive force of a multimodal study to a solitary fMRI classifier and distinguishes construct convergence from mere overlap of reported anatomical labels. Poldrack (2006) Kriegeskorte et al. (2008) Nichols et al. (2017)

Recent Empirical Directions: Autobiographical Imagery and Semantic Generalization

Recent investigations extend decoding beyond repeated laboratory labels by testing whether previously learned semantic mappings transfer to more naturalistic psychological activity. Anderson and colleagues examined fMRI data from fifty participants who imagined autobiographical versions of twenty common scenarios, and applied a semantic decoder learned from sentence-reading data collected in a separate participant group. The investigators assessed reconstruction of participant-specific experiential feature ratings and reported above-chance reconstruction for a subset of those features. This study is important because its research question concerns transfer between participant groups and between distinct cognitive tasks, rather than fitting a new decoder directly to each autobiographical imagination dataset. Anderson et al. (2025)

Transfer is an exacting test, and its interpretation depends on the defined targets. The relevant outcome in Anderson and colleagues' study was the reconstruction of numerical feature ratings associated with prompted scenarios, not unrestricted recovery of a complete personal memory. Written scenario cues, semantic similarity, large-scale spatial averaging, and the chosen feature vocabulary remain part of the experimental situation and of the interpretation. Their report explicitly examined whether reconstruction reflected individual characteristics beyond generic prompt content, furnishing a methodological example of how a favorable decoder should be challenged with a plausible alternative explanation. The resulting evidence supports transferable feature-level correspondence between some language and autobiographical-imagery representations, with remaining uncertainty about neural implementation and generalization outside the examined scenarios. Anderson et al. (2025) Poldrack (2006)

A complementary 2026 synthesis by Radman and colleagues employed activation likelihood estimation across prior fMRI studies to examine semantic-category representations. Such coordinate-based meta-analytic methods investigate convergence among reported activation locations, whereas trial-wise MVPA tests predictive information in individual datasets. They are complementary forms of evidence, and a spatial convergence map does not by itself establish multivariate cross-subject decoding. Methodological integration should ask whether a priori regions suggested by independent synthesis also provide information under controlled prediction tests and whether results survive a switch from task-specific labels to semantically matched but untrained conditions. This illustrates the ongoing transition from proof-of-concept classification toward transferable accounts of cortical representational structure. Radman et al. (2026) Kriegeskorte et al. (2006) Nichols et al. (2017)

Clinical and Translational Interpretation

Clinical enthusiasm for neuroimaging prediction rests on an understandable aspiration: a biological measurement might contribute information not captured by interview, behavior, or routine assessment. Yet a classifier trained to distinguish diagnosed patients from controls may exploit sampling, medication status, movement, comorbidity, scanner site, and acquisition differences. Its accuracy for a carefully balanced research comparison does not automatically translate to predictive value in a real clinical population with different prevalence and diagnostic uncertainty. Clinical validation therefore requires a prespecified intended use, representative samples, explicit benchmarks against existing assessments, calibration and uncertainty reporting, and evaluation under external settings. Population associations documented by large imaging studies reinforce the need to plan sample size around expected effects, not around the computational sophistication of the decoder. Marek et al. (2022) Varoquaux (2018) Poldrack et al. (2017)

A translational system must distinguish retrospective description from prospective decision support. If the target is future outcome or response to intervention, model training must precede the observations that define the outcome, and evaluation must respect temporal order. Participant identity, family structure, site, and repeated visits must not leak across development and test partitions in ways that render a prediction deceptively easy. Prevalence, cost of errors, fairness across demographic and clinical subgroups, and meaningful incremental benefit become primary endpoints. A high overall area-under-curve statistic cannot substitute for understanding what happens to individuals exposed to false positives or false negatives. Such considerations are especially salient when the imaging instrument is expensive and the claimed application is screening or differential diagnosis. Marek et al. (2022) Allefeld et al. (2016) Nichols et al. (2017)

Clinical fMRI is also vulnerable to altered hemodynamics. Neurological illness, development, aging, medication, cerebrovascular disease, and individual vascular reactivity can modify the mapping between neural activity and BOLD signal. A decoding model could therefore confuse a change in measurement physiology with a change in psychological function. Assessment must consider vascular and motion-related confounds, spatial coverage, test–retest performance, and complementary clinical evidence. This monograph is methodological and does not provide individualized diagnostic or therapeutic recommendations. Independent prospective evaluation and qualified clinical review would be necessary before any claimed medical intervention or diagnostic decision built upon its conclusions. Logothetis et al. (2001) Buxton et al. (1998) Nichols et al. (2017)

Integrated Evaluation Framework and Open Research Problems

A comprehensive evaluation of fMRI-based decoding can be organized along five connected questions: whether the raw measurement reflects the intended physiology, whether the task operationalizes the proposed psychological construct, whether the estimator predicts genuinely independent observations, whether the result is robust to plausible alternative analytic decisions, and whether its claimed application generalizes to the population and setting of interest. These questions follow a causal chain from acquisition to interpretation, but each has its own failure modes. A statistically strong decoder can coexist with weak psychological specificity; robust neural effects can coexist with small clinical utility; and anatomically replicable patterns can remain ambiguous about the microscopic code that generated them. Logothetis et al. (2001) Kriegeskorte et al. (2009) Marek et al. (2022)

Experiments designed for explanatory progress should combine preregistered contrasts, independent regions, nested parameter selection, dependence-respecting cross-validation, permutation tests with valid exchangeability, and transparent reporting of effect sizes and individual distributions. Competing feature models should be evaluated on the same held-out items, and transfer should be tested along named dimensions such as novel stimuli, different task instructions, independent participants, sites, and time periods. A useful validation table specifies the training domain, test domain, target, participant count, independent unit, nuisance controls, primary metric, uncertainty procedure, and planned failure criteria. Public release of deidentified data and reproducible analysis pipelines, wherever ethically appropriate, permits other teams to audit and extend those conclusions. Varoquaux (2018) Nichols et al. (2017) Poldrack et al. (2017)

Open problems include disentangling microscopic neural coding from coarse vascular sampling, identifying when decoders track perception rather than attentional or motor consequences, quantifying the reliability of single-trial and naturalistic predictions, and transferring representations across heterogeneous people and scanners without losing psychological meaning. Generative reconstructions require metrics that can distinguish brain-derived constraints from image or language priors. Clinical translation requires prospective evidence of benefit over accessible behavioral assessments. The availability of increasingly capable artificial models intensifies the need for strong controls because a fluent output may obscure a weak physiological constraint. Progress should be measured by the precision of its tested hypotheses, the strength of its counterfactual comparisons, and transparent performance under conditions not optimized to yield impressive demonstrations. Takagi & Nishimoto (2023) Tang et al. (2023) Anderson et al. (2025)

Conclusion

Functional MRI and multivariate neural decoding have established that human BOLD patterns contain experimentally useful information about sensory distinctions, category structure, attention, task variables, semantic features, and selected aspects of naturalistic cognition. The historical transition from oxygenation-dependent MRI contrast to distributed classification, encoding models, representational similarity analysis, and generative reconstruction has increased the scope of questions available to cognitive neuroscience. Its scientific value is greatest when researchers specify the target variable and demonstrate that the predictive relation survives appropriate controls and independent evaluation. Neurovascular mediation, anatomical variability, preprocessing, statistical selection, and model priors remain constitutive conditions of every inference drawn from these measurements. Ogawa et al. (1990) Haxby et al. (2001) Kriegeskorte et al. (2008) Tang et al. (2023)

The principal lesson of the evidence is that informative measurement, psychological interpretation, and causal explanation are distinct but compatible research achievements. Reliable discrimination of a target is evidence of accessible information; comparison of independently validated representational models can narrow the explanatory possibilities; convergent physiological and perturbational studies can further test proposed mechanisms. This progression makes the field scientifically cumulative. Claims of generalized brain reading or clinical precision become assessable through their actual training requirements, task constraints, external performance, measurement reliability, and errors. The next advances will depend on stronger multimodal explanations, large and diverse samples where population-level claims are made, and transparent evidence of transfer across contexts. Poldrack (2006) Logothetis et al. (2001) Marek et al. (2022)

The research synthesis therefore supports a disciplined formulation: multivariate fMRI decoding is a family of conditional predictive measurements through which investigators can test hypotheses about the organization of human cognitive processes. A decoded pattern does not settle what caused a mental state or exhaust the state itself; the science develops by making the mapping between experimental constructs and measured responses increasingly explicit, testable, and generalizable. Within the broader Artificial Era, this same standard permits careful investigation of what artificial models recover from biological observations and how human cognition can be studied through their partnership. The future of the method lies in higher evidential resolution: better-defined constructs, improved physiological control, stronger external validation, and precise attribution of what prediction establishes. Naselaris et al. (2011) Hebart & Baker (2018) Anderson et al. (2025)

References

  • Allefeld, C., Görgen, K., & Haynes, J.-D. (2016). Valid population inference for information-based imaging: From the second-level t-test to prevalence inference. NeuroImage, 141, 378–392. https://doi.org/10.1016/j.neuroimage.2016.07.040
  • Anderson, A. J., Fernandino, L., Gross, W. L., Rey, H. G., & Binder, J. R. (2025). Neural decoding of autobiographical mental image features with a general semantic model. Nature Communications, 16, Article 11208. https://doi.org/10.1038/s41467-025-65541-1
  • Belliveau, J. W., et al. (1991). Functional mapping of the human visual cortex by magnetic resonance imaging. Science, 254(5032), 716–719. https://doi.org/10.1126/science.1948051
  • Botvinik-Nezer, R., et al. (2020). Variability in the analysis of a single neuroimaging dataset by many teams. Nature, 582, 84–88. https://doi.org/10.1038/s41586-020-2314-9
  • Boynton, G. M., Engel, S. A., Glover, G. H., & Heeger, D. J. (1996). Linear systems analysis of functional magnetic resonance imaging in human V1. The Journal of Neuroscience, 16(13), 4207–4221. https://doi.org/10.1523/JNEUROSCI.16-13-04207.1996
  • Buxton, R. B., Wong, E. C., & Frank, L. R. (1998). Dynamics of blood flow and oxygenation changes during brain activation: The balloon model. Magnetic Resonance in Medicine, 39(6), 855–864. https://doi.org/10.1002/mrm.1910390602
  • Cox, D. D., & Savoy, R. L. (2003). Functional magnetic resonance imaging (fMRI) “brain reading”: Detecting and classifying distributed patterns of fMRI activity in human visual cortex. NeuroImage, 19(2), 261–270. https://doi.org/10.1016/S1053-8119(03)00049-1
  • Davis, T., LaRocque, K. F., Mumford, J. A., Norman, K. A., Wagner, A. D., & Poldrack, R. A. (2014). What do differences between multi-voxel and univariate analysis mean? How subject-, voxel-, and trial-level variance impact fMRI analysis. NeuroImage, 97, 271–283. https://doi.org/10.1016/j.neuroimage.2014.04.037
  • Esteban, O., et al. (2019). fMRIPrep: A robust preprocessing pipeline for functional MRI. Nature Methods, 16, 111–116. https://doi.org/10.1038/s41592-018-0235-4
  • Gorgolewski, K. J., et al. (2016). The brain imaging data structure, a format for organizing and describing outputs of neuroimaging experiments. Scientific Data, 3, Article 160044. https://doi.org/10.1038/sdata.2016.44
  • 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
  • Haxby, J. V., Guntupalli, J. S., Connolly, A. C., Halchenko, Y. O., Conroy, B. R., Gobbini, M. I., Hanke, M., & Ramadge, P. J. (2011). A common, high-dimensional model of the representational space in human ventral temporal cortex. Neuron, 72(2), 404–416. https://doi.org/10.1016/j.neuron.2011.08.026
  • Haynes, J.-D., & Rees, G. (2006). Decoding mental states from brain activity in humans. Nature Reviews Neuroscience, 7, 523–534. https://doi.org/10.1038/nrn1931
  • Hebart, M. N., & Baker, C. I. (2018). Deconstructing multivariate decoding for the study of brain function. NeuroImage, 180(Part A), 4–18. https://doi.org/10.1016/j.neuroimage.2017.08.005
  • Kamitani, Y., & Tong, F. (2005). Decoding the visual and subjective contents of the human brain. Nature Neuroscience, 8, 679–685. https://doi.org/10.1038/nn1444
  • Kamitani, Y., & Tong, F. (2006). Decoding seen and attended motion directions from activity in the human visual cortex. Current Biology, 16(11), 1096–1102. https://doi.org/10.1016/j.cub.2006.04.003
  • Kay, K. N., Naselaris, T., Prenger, R. J., & Gallant, J. L. (2008). Identifying natural images from human brain activity. Nature, 452, 352–355. https://doi.org/10.1038/nature06713
  • Kriegeskorte, N., Goebel, R., & Bandettini, P. (2006). Information-based functional brain mapping. Proceedings of the National Academy of Sciences, 103(10), 3863–3868. https://doi.org/10.1073/pnas.0600244103
  • Kriegeskorte, N., Mur, M., & Bandettini, P. (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., Simmons, W. K., Bellgowan, P. S. F., & Baker, C. I. (2009). Circular analysis in systems neuroscience: The dangers of double dipping. Nature Neuroscience, 12, 535–540. https://doi.org/10.1038/nn.2303
  • Kwong, K. K., et al. (1992). Dynamic magnetic resonance imaging of human brain activity during primary sensory stimulation. Proceedings of the National Academy of Sciences, 89(12), 5675–5679. https://doi.org/10.1073/pnas.89.12.5675
  • 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
  • Marek, S., et al. (2022). Reproducible brain-wide association studies require thousands of individuals. Nature, 603, 654–660. https://doi.org/10.1038/s41586-022-04492-9
  • Misaki, M., Kim, Y., Bandettini, P. A., & Kriegeskorte, N. (2010). Comparison of multivariate classifiers and response normalizations for pattern-information fMRI. NeuroImage, 53(1), 103–118. https://doi.org/10.1016/j.neuroimage.2010.05.051
  • Naselaris, T., Kay, K. N., Nishimoto, S., & Gallant, J. L. (2011). Encoding and decoding in fMRI. NeuroImage, 56(2), 400–410. https://doi.org/10.1016/j.neuroimage.2010.07.073
  • Nichols, T. E., et al. (2017). Best practices in data analysis and sharing in neuroimaging using MRI. Nature Neuroscience, 20, 299–303. https://doi.org/10.1038/nn.4500
  • Norman, K. A., Polyn, S. M., Detre, G. J., & Haxby, J. V. (2006). Beyond mind-reading: Multi-voxel pattern analysis of fMRI data. Trends in Cognitive Sciences, 10(9), 424–430. https://doi.org/10.1016/j.tics.2006.07.005
  • Ogawa, S., Lee, T. M., Kay, A. R., & Tank, D. W. (1990). Brain magnetic resonance imaging with contrast dependent on blood oxygenation. Proceedings of the National Academy of Sciences, 87(24), 9868–9872. https://doi.org/10.1073/pnas.87.24.9868
  • 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
  • Poldrack, R. A., et al. (2017). Scanning the horizon: Towards transparent and reproducible neuroimaging research. Nature Reviews Neuroscience, 18, 115–126. https://doi.org/10.1038/nrn.2016.167
  • Radman, M., Podmore, J. J., Poli, R., Paulmann, S., & Daly, I. (2026). Decoding semantic categories: Insights from an fMRI ALE meta analysis. Journal of Neural Engineering, 22(6), Article 061006. https://doi.org/10.1088/1741-2552/ae302b
  • Takagi, Y., & Nishimoto, S. (2023). High-resolution image reconstruction with latent diffusion models from human brain activity. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 14453–14463. https://openaccess.thecvf.com/content/CVPR2023/html/Takagi_High-Resolution_Image_Reconstruction_With_Latent_Diffusion_Models_From_Human_Brain_CVPR_2023_paper.html
  • Tang, J., LeBel, A., Jain, S., & Huth, A. G. (2023). Semantic reconstruction of continuous language from non-invasive brain recordings. Nature Neuroscience, 26, 858–866. https://doi.org/10.1038/s41593-023-01304-9
  • Varoquaux, G. (2018). Cross-validation failure: Small sample sizes lead to large error bars. NeuroImage, 180(Part A), 68–77. https://doi.org/10.1016/j.neuroimage.2017.06.061
  • Wang, S., Dubarry, A.-S., Chanoine, V., Sein, J., Anton, J.-L., Nazarian, B., Mercier, M. R., Trébuchon, A., & Pattamadilok, C. (2025). Revealing the co-existence of written and spoken language coding neural populations in the visual word form area. Imaging Neuroscience, 3, Article imag_a_00524. https://doi.org/10.1162/imag_a_00524
  • Woolrich, M. W., Ripley, B. D., Brady, M., & Smith, S. M. (2001). Temporal autocorrelation in univariate linear modeling of FMRI data. NeuroImage, 14(6), 1370–1386. https://doi.org/10.1006/nimg.2001.0931
  • Yarkoni, T., Poldrack, R. A., Nichols, T. E., Van Essen, D. C., & Wager, T. D. (2011). Large-scale automated synthesis of human functional neuroimaging data. Nature Methods, 8, 665–670. https://doi.org/10.1038/nmeth.1635