Five tones form a pattern. The fifth is different from the first four, and that difference is familiar enough to become part of the pattern itself. When a rare sequence arrives in which the fifth tone is the same, a listener encounters two kinds of regularity at once: a local relation among tones and a broader relation among whole sequences. A scalp response to the unusual sequence may therefore be informative about how the brain handles structure at different levels. It is not, by itself, a photograph of predictions travelling down a neural hierarchy.
Measuring hierarchical predictive coding means testing whether observations distinguish predictions and errors at more than one specified level of a generative model. The word hierarchical does important work. A neural response to any unexpected stimulus may support a broad account of prediction. A hierarchy claim additionally requires evidence about distinct representational levels, their dependence on one another and, for a stronger mechanistic claim, directed communication between them. This Methods & Measures entry asks how tasks and recordings can support those different claims without treating them as equivalent.
What exactly would be measured?
A hierarchy is not simply a collection of brain regions that light up at different times. It is an organized account in which one level represents a regularity, another level represents a broader regularity or hidden cause, and information at one level constrains inference at another. In the auditory example, a local level might concern the probability of one tone after another. A global level might concern the probability of a five-tone sequence among other sequences. These levels are defined by the task and model, not automatically by anatomical labels.
Predictive coding theories propose that predictions can flow from higher to lower levels while mismatches can influence higher levels. However, an experiment usually measures electrical activity, magnetic fields, blood oxygenation or behavior, not a labelled message saying “prediction” or “error.” A response difference becomes evidence for a proposed computation only through a design that makes one interpretation more plausible than rivals. The same observed signal may be consistent with novelty detection, adaptation, attention or task demands.
Construct validity therefore has several layers of its own. The stimulus design must create distinguishable local and global expectations. The participant or animal must have an opportunity to learn them. The measurement must be sensitive to their consequences. The analysis must compare rival explanations. Finally, the conclusion must specify whether the evidence concerns behavioral sensitivity, spatially distinct responses, temporal ordering, layer-dependent activity or effective connectivity. These outcomes differ in what they establish.
Wacongne and colleagues used auditory sequences to manipulate regularities at two levels and recorded human electrical and magnetic responses. Their work is an important example of a task in which a locally deviant tone can be globally expected. It supports the possibility that responses to local violations are modulated by broader sequence expectations. The finding is more informative than a simple rare-tone response because the design gives researchers two regularities to compare. It still depends on the precise stimulus frequencies, learning conditions and interpretation of scalp signals.
Local and global tasks
A local-global paradigm typically constructs short sequences and varies both the relation between items within a sequence and the frequency of sequence types across a block. One five-tone sequence might end in a deviant tone yet occur frequently. Another might contain five identical tones yet occur rarely. The first has local deviation without global rarity; the second may have global rarity without local deviation. This crossing creates a sharper question than asking whether an infrequent tone produces a larger response.
The design is not a perfect two-by-two switch inside the brain. Participants may count tones, attend to the fifth position, notice block changes or learn only part of the structure. The globally rare sequence is still rare in its recent presentation history. Its physical sequence may differ from a comparator. Analyses must therefore specify which contrast controls physical input, recent repetition and attention as well as possible. A claim about global prediction should name the blockwise rule and show that the effect is not simply a response to a physically new sequence.
Uhrig, Dehaene and Jarraya used local and global auditory regularities in awake macaques. Their fMRI results distinguished responses to first-order violations in auditory cortex from responses to second-order violations in a frontoparietal network. That spatial distinction supports hierarchical organization of auditory novelty responses in that species and task. It does not show that every region has a unique computational level, nor does fMRI alone reveal which region sent a prediction to which other region.
Jiang and colleagues combined high-field fMRI with surface electrocorticography in awake marmosets during local-global auditory sequences. They reported shorter-timescale local processing along earlier auditory pathways and longer-timescale global sequence processing in anterior temporal and prefrontal areas. The combination of spatial and temporal information is valuable: two modalities can constrain an interpretation more than either one alone. Nevertheless, the task remains an operational model of selected auditory regularities, not a complete map of the animal’s internal beliefs.
One useful audit question is whether the study manipulates the level of regularity or only labels it after observing a response. If a design varies one tone’s probability but not a rule over sequences, it cannot by itself test global sequence prediction. If it varies whole-sequence frequency but leaves local item probability confounded with it, the levels cannot be cleanly separated. A robust hierarchy claim starts with separable, preregistered contrasts or a clearly stated model that yields different predictions for the levels.
Time and anatomy are clues, not labels
Early and late responses can suggest a processing sequence. In auditory experiments, a relatively early response to a local deviance and a later response to a global deviance fit an account in which different regularities are processed at different stages. Yet latency alone does not assign a computational level. The later response may require more evidence accumulation, greater attention, a decision or conscious report. It could be late because the task is harder, not because a message passed through a higher predictive level.
Anatomy adds another kind of constraint. If distinct contrasts recruit auditory cortex and frontoparietal areas differently, a single undifferentiated novelty mechanism becomes less adequate. But mapping “lower” onto one anatomical site and “higher” onto another is an inference from functional organization, not a direct reading of an internal hierarchy. A region may participate in several computations; regions can contain mixed neural populations; and spatially broad fMRI signals pool across many cells and vessels.
Laminar imaging attempts to add depth within cortex. A recent human auditory study used high-resolution blood-oxygen imaging with a cascading oddball task and modeled vascular drainage. It found patterns the authors interpreted as potentially consistent with prediction-violation signals in superficial layers of a lower-level auditory region, middle-layer modulation in a higher-order temporal area and deeper-layer effects associated with updating. The authors’ own language is appropriately cautious. Layer-sensitive BOLD signals can constrain a circuit hypothesis, but they do not identify cell types or show an individual neural message moving between layers.
Vascular effects deserve particular attention. A blood-oxygen signal measured near the cortical surface may partly reflect blood draining from deeper tissue. Without acquisition choices and explicit modelling of this effect, an apparent layer profile may not be a neuronal layer profile. Even with corrections, uncertainty remains because the model of vascular contamination has assumptions. Laminar results become stronger when converging evidence from other modalities and tasks supports the same directional account.
Cross-species convergence is useful but bounded. A human scalp response, macaque fMRI contrast and marmoset ECoG pattern may all fit a hierarchical interpretation. They do not form a single replication because species, tasks, arousal states and measurements differ. Their value is triangulation: each method has limitations that differ from the others. The synthesis should state the shared pattern and the parts that remain method-specific.
Connectivity models and the inference problem
A theory about ascending errors and descending predictions is a theory about interactions. Dynamic causal modelling offers one way to compare candidate network architectures against measured evoked responses. Garrido and colleagues modelled auditory responses to frequency deviants and assessed alternative accounts involving changes in connections and local neuronal responses. The method can make a mechanistic hypothesis explicit. Its output is conditional on the set of candidate models, priors, data quality and observation model.
A winning connectivity model is not direct proof that the brain uses exactly those connections for prediction messages. If a plausible rival model was omitted, model comparison cannot reject it. Different combinations of parameters can sometimes produce similar measured signals. The scalp recording itself is an indirect mixture of sources. A transparent report should list the model space, justify the nodes and priors, evaluate model recovery where possible and show whether conclusions survive reasonable alternatives.
The same caution applies to claims about direction. A temporal lead in one region followed by activity in another does not establish a causal signal travelling between them. Common input, recurrent loops and measurement delays can produce apparent order. Anatomical tracing, perturbation or more specific recordings can strengthen directionality, but each brings its own limits. The measured result should be described in the language of the method: for example, a model favored one connectivity pattern under a defined comparison, rather than an unconditional demonstration of top-down prediction.
There is also a conceptual distinction between a hierarchy of regularity representations and a specific predictive-coding algorithm. A system could detect local and global anomalies using learned templates, recurrent memory, attention or another computation. Showing two response levels supports hierarchical processing. Showing that the levels communicate through a particular prediction-error scheme requires more discriminating tests. That distinction makes the field more precise, not less interesting.
Separating expectation from repetition
Repeated events often produce smaller neural responses. Predictive coding can explain some of this reduction as a consequence of fulfilled expectations. Sensory adaptation can also reduce responses because recently active neural populations change their sensitivity. A conventional oddball sequence confounds the two: standards are repeated and expected, while deviants are rare and surprising. Subtracting one response from the other does not isolate a pure error term.
Controls can present the same physical tone in contexts with different expectations while matching recent exposure as carefully as the task allows. Roving sequences can change which tone is standard. Omission trials remove a predicted event, reducing the confound of a physically novel deviant, although an omission response can still depend on timing, attention and the exact sequence. A cascade or many-standards control can help test whether a deviant response exceeds what rarity and adaptation predict. No single control settles every mechanism.
Tang and colleagues used human EEG and multivariate modelling to separate expectation and repetition effects for visual gratings. Unexpected events enhanced an orientation-selective representation, whereas repetition did not show the same effect on selectivity. This result is not an auditory hierarchy test. It is a valuable methodological warning: neural response amplitude, representational selectivity, repetition and expectation need not move together. A study that only compares average amplitude may miss that distinction.
Attention is another rival and possible partner. A participant instructed to find a rare global pattern may allocate resources differently from someone who passively hears tones. If local and global effects change with attentional state, researchers should report that dependence rather than subtract attention away conceptually. Attention can alter the precision or relevance of incoming information in some predictive models, yet the task still needs empirical evidence for that account. A larger late response under active counting is not automatically a stronger high-level prediction error.
Reliability, model checks and individual differences
Many studies establish average effects across participants or trials. An individual-level measure requires a separate validity argument. A group-average local-global contrast may be robust even when a person’s score changes with stimulus set, fatigue, head movement, electrode placement or analysis choices. Before interpreting a score as a stable trait, researchers need test-retest evidence, measurement uncertainty and clarity about the population in which the score is intended to be used.
The task should also demonstrate learning. If participants fail to acquire the global regularity, a null global effect cannot be read as evidence that their brain lacks hierarchical processing. A behavioral report, occasional probe or independent measure can help establish that the pattern was available to the participant, although probing can itself change attention. Researchers should explain that tradeoff. For nonhuman animals, behavioral training and physiology can provide different constraints, but assumptions about what was learned still need testing.
Computational models can generate trialwise predictions for each level. Their value depends on whether they predict unseen responses better than alternatives and whether fitted parameters can be recovered from data of the available quality. A complex model may fit recorded signals while remaining too flexible to identify a unique mechanism. Researchers should compare simpler memory or adaptation models, examine parameter uncertainty and avoid presenting a fitted latent state as though it were directly observed.
Analysis flexibility creates another threat. A hierarchy can seem clearer when time windows, channels, cortical regions or frequency bands are selected after seeing which ones show the desired contrast. Independent replication, held-out data, preregistration and transparent reporting help readers distinguish a predicted pattern from an exploratory one. These practices are especially important when a study combines several modalities and many plausible choices of preprocessing and model specification.
Reading a hierarchy claim carefully
First ask what the levels are. Are they tone features, short sequences, long sequences, spatial scales, abstract rules or cortical layers? The word “higher” should have a definition before a brain map appears. Then ask whether the task manipulates those levels independently enough to compare them. A response to a rare event is not enough; the design must test different kinds of rarity or different predictions.
Next identify the measured quantity. Is it a scalp potential, magnetic field, ECoG power, BOLD contrast, behavior or a parameter fitted by a model? Each is an observation at a different scale. A predicted-error unit in a mathematical model cannot be equated with a voltage difference without an explicit observation model. The more direct the claim sounds, the more important this translation becomes.
Look for the competing account. Could local effects reflect adaptation? Could global effects reflect counting or surprise about block frequency? Could an apparent layer signal reflect vascular drainage? Could a connectivity result depend on which models were compared? An article that names and tests these alternatives contributes more to construct validity than one that merely repeats a favored theoretical vocabulary.
Finally, scale the conclusion. “The response distinguished local and global auditory violations in this paradigm” is a measured result. “These results are consistent with a hierarchical predictive account” is an interpretation. “We have identified a universal cortical prediction-error circuit” is a much larger claim requiring far more evidence. The Institute’s checklist here is editorial synthesis of methods concerns, not a validated diagnostic instrument.
What the evidence does not show
The cited studies do not provide a single direct readout of predictions and errors at every level of cortex. They do not establish that all novelty responses are prediction errors, that local and global effects necessarily use the same algorithm, or that a later or more frontal response is automatically a higher-level prediction. They do not justify clinical, educational or workplace classification of individuals from these experimental measures.
Spatial gradients, latency differences, layer-sensitive signals and connectivity-model preferences each constrain theories differently. None alone proves the full direction and content of neural messages proposed by hierarchical predictive coding. Adaptation, physical stimulus differences, attention, conscious report, vascular effects and model-space restrictions remain material alternatives or moderators. Findings in marmosets and macaques cannot simply be transferred to humans without examining the task and species boundaries.
The research supports a disciplined program of hypothesis testing: define levels, create contrasts that distinguish them, use multiple measures where feasible and make competing explanations explicit. It does not turn a broad theory into a universal biomarker. The strongest contribution of a hierarchy experiment is often a narrower one: it shows which regularities a nervous system distinguishes under specified conditions, and which mechanistic questions remain open.
Back to the five tones
The final tone is not “surprising” in only one sense. Its pitch may differ locally while the complete sequence is familiar globally, or its pitch may match while the complete sequence is rare. By crossing those possibilities, the experiment asks a question about levels of regularity rather than novelty alone. A measured difference can be meaningful even when it cannot reveal every hidden message in the circuit.
Hierarchical predictive coding becomes testable when “hierarchy” is operational, not decorative. The credible claim identifies the regularities, the observations, the controls and the inference connecting them. That discipline allows a useful result to remain useful even if later work revises the mechanism behind it.
Research references
- Wacongne, C., Labyt, E., van Wassenhove, V., Bekinschtein, T., Naccache, L., & Dehaene, S. (2011). Evidence for a hierarchy of predictions and prediction errors in human cortex. Proceedings of the National Academy of Sciences, 108(51), 20754–20759. View source
- Uhrig, L., Dehaene, S., & Jarraya, B. (2014). A hierarchy of responses to auditory regularities in the macaque brain. Journal of Neuroscience, 34(4), 1127–1132. View source
- Jiang, Y., et al. (2022). Constructing the hierarchy of predictive auditory sequences in the marmoset brain. eLife, 11, e74653. View source
- Garrido, M. I., Kilner, J. M., Kiebel, S. J., Stephan, K. E., & Friston, K. J. (2009). Dynamic causal modeling of the response to frequency deviants. Journal of Neurophysiology, 101(5), 2620–2631. View source
- Tang, M. F., et al. (2018). Prediction error and repetition suppression have distinct effects on neural representations of visual information. eLife, 7, e33123. View source
- Faes, L. K., et al. (2026). Predictive acoustical processing in human cortical layers. Nature Communications, 17, 6862. View source