HED vs CogPO¶
HED, the Hierarchical Event Descriptors, and CogPO, the Cognitive Paradigm Ontology, both describe what happens to a participant in an experiment in a controlled vocabulary, and they do it at different levels: CogPO labels a condition by its stimuli, responses and instructions, as the experimenter planned it; HED labels each event as it happened. The two are related by history as well as by design. This page sets them side by side: where they came from, what each describes, where their vocabularies meet, what each has that the other lacks, and how the Catalog fits between them.
The account rests on the two design papers: Turner and Laird (2012), “The Cognitive Paradigm Ontology: Design and Application”, Neuroinformatics 10:57-66 (DOI), and Robbins, Truong, Jones, Callanan and Makeig (2022), “Building FAIR functionality: Annotating events in time series data using Hierarchical Event Descriptors (HED)”, Neuroinformatics 20:463-481 (DOI).
Two designs for one problem¶
Both projects began from the same complaint. Turner and Laird open with the observation that a paradigm named in the literature, “the Stroop paradigm or the Sternberg paradigm”, can “vary tremendously in the stimuli that are presented to the subject, the response expected from the subject, and the instructions given to the subject”, so that named tasks cannot be compared across studies. Robbins and colleagues open with the same gap seen from the data side: shared recordings rarely say what their participants did in a form a machine can act on, which blocks analysis across studies.
The two answers differ in what they take as the unit of description.
CogPO describes the plan. It grew out of the BrainMap database’s taxonomy for coding published neuroimaging experiments, and it keeps BrainMap’s stance: the ontology represents “what was intended by the experimenters and what instructions were given, and not what the subject may have actually perceived about the stimuli or what they thought they were doing”. Its unit is the condition, “a planned process which must include at least one stimulus, one instruction, and one response type”, and a paradigm class is a recurring pattern of conditions, named once BrainMap had seen “at least five instances” of it. CogPO is an OWL ontology on the Basic Formal Ontology with IAO annotation properties, built for curators coding papers into a database and for reasoning over the result. Its authors expected its Stimulus and Response classes to become “inferred classes” once ontologies of sounds, shapes, food and actions existed to point at.
HED describes what happened. It was first deployed in 2011 to annotate the events in EEG recordings shared through the HeadIT repository, and its unit has always been the event: a time-stamped occurrence during a recording, described by a string of tags. HED is a tag vocabulary with a grammar for combining tags into groups and definitions, a validator, and tools that read the annotations back for analysis; it is built for the people who produce data, not for curators of the literature, and it was adopted by BIDS in 2019. Where CogPO stops at the instruction “push a button when the light flashes”, HED records each flash and each press, including the presses that did not come.
The connection. Early HED-1G terminology, the 2022 paper says, “was partially based
on CogPO”. Second-generation HED carried a Paradigm hierarchy among its top-level terms,
and the plan through HED-1G and HED-2G was “to incorporate the CogPO list of task
paradigms, with hopes of linking HED event descriptions to task databases such as the
Cognitive Atlas”. Third-generation HED removed the Paradigm tags, “because the available
paradigm nomenclature is not standardized”: task descriptions in the Cognitive Atlas
“vary in specificity and use only broadly-defined and sub-field specific terminology” and
“do not, at present, represent machine-actionable information”. What remains in the HED
schema is a slot for the identifier, Metadata-identifier/CogPo/#, beside one for the
Cognitive Atlas, and the paper is plain that “merely associating these respective IDs
with the recording does not present task information in a machine-actionable form”. The
same section names the missing piece: “a task specification meta-code to enable
comparison of events on the basis of their function and value in the context of the
task”. That is the gap the Catalog exists to fill.
What each describes¶
A CogPO paradigm class is defined by its conditions, and a condition is “a planned combination of stimuli and instructions regarding responses”. CogPO’s one worked example in the OWL file is an FBIRN auditory oddball task with two conditions, standard tones and target tones; each condition has a stimulus (tones, auditory modality, in the role of standard or target), an instruction (detect) and a response (button press with the hand). Nothing in CogPO says when a tone was played or which one.
HED describes the same experiment as a sequence of events. Each tone onset is a
Sensory-event with Auditory-presentation, Tone and a role tag (Target or
Non-target), each press an Agent-action with Press and Participant-response, and
each carries a time. The condition CogPO names is recovered from the events by their
shared tags, and HED’s Condition-variable and Definition mechanisms name it
explicitly when the annotator wants to. Where CogPO stops at “in at least one of the
conditions, subjects detect an oddball tone”, HED records every tone and every press
that made up the run.
The two therefore meet at the condition. CogPO’s condition is a summary of a HED event
stream; a HED event stream is an instance of a CogPO condition unfolding in time. And the
deviations CogPO’s authors said their formulation could not represent, “the subject
failed to press the button, for example, or pressed the button more than was
instructed”, are ordinary HED annotations: an Omitted-action, an extra
Participant-response.
Where the vocabularies meet¶
CogPO’s six dimension vocabularies each land in one HED subtree, which is why the Catalog could adopt them as facets and pair every value with a tag:
CogPO dimension |
HED 8.4.0 subtree |
|---|---|
Stimulus Modality |
|
Explicit Stimulus |
|
Stimulus Role |
|
Response Modality |
|
Overt Response |
|
Instructions |
|
Of the 102 CogPO dimension values, 88 have a HED 8.4.0 tag and 58 of those are an exact tag, the same concept under the same or an equivalent name; the rest have a nearest parent tag or none.
Facet |
Values |
From CogPO |
Added from HED |
Added by the Catalog |
CogPO values with a HED tag |
of which an exact tag |
|---|---|---|---|---|---|---|
Stimulus modality |
10 |
10 |
0 |
0 |
9 |
5 |
Stimulus kind |
46 |
41 |
0 |
5 |
31 |
12 |
Stimulus role |
12 |
4 |
8 |
0 |
4 |
4 |
Response modality |
11 |
10 |
0 |
1 |
10 |
8 |
Response kind |
23 |
18 |
0 |
5 |
15 |
11 |
Instructions |
32 |
19 |
8 |
5 |
19 |
18 |
The value-by-value tables, with the HED tag and a note on the fit for every value, are
on the dimension mapping page. The fit is close for a
reason: Turner and Laird expected exactly this, writing that “tones are not really a
stimulus type, but instead they are a kind of sound” and that CogPO would one day refer
to “ontologies of sounds, shapes, food, actions etc.” HED’s Item and Action trees are
such vocabularies, and CogPO’s stimulus and response classes map onto them as the roles
its authors said they were.
What HED has that CogPO does not¶
Time. Every HED event has an onset, and may have a duration or an offset. CogPO has no temporal vocabulary at all: a condition is a set, not a sequence.
Structure. HED’s Event subtree distinguishes sensory events, agent actions, data
features, experiment control and experiment structure; its organizational tags
(Experimental-trial, Time-block, Condition-variable, Control-variable,
Definition) let an annotator say which events belong to which trial, block and
condition. CogPO’s structure is the paradigm-condition-component hierarchy and nothing
below it.
What actually happened. CogPO records the intended paradigm and, by its authors’
account, cannot record a missed or extra response. HED records the response that was
made, or not made, and its Task-action-type tags (Correct-action, Incorrect-action,
Omitted-action, Miss) say how it related to what was asked.
Richer roles. CogPO has four stimulus roles, and those only on its wiki. HED’s
Task-stimulus-role has 22 (Distractor, Cue, Oddball, Novel, Penalty,
Go-signal, Stop-signal, Priming, Threat and more) and Task-event-role adds
Cue, Feedback, Instructional, Participant-response, Warning. The Catalog’s
facet vocabulary takes eight roles from HED that CogPO never had.
More actions. HED’s Action/Think has 21 verbs to CogPO’s 17 instructions, and
several the Catalog’s tasks need: Judge, Recognize, Predict, Plan, Learn,
Identify, Switch-attention. HED also names the agent (Experiment-participant,
Experimenter) and the agent’s state (Eyes-closed, Resting, Passive), which CogPO
leaves implicit.
A grammar and tools. HED tags combine into groups and definitions with a validator behind them; a HED string can be checked and searched, and library schemas extend the vocabulary for a field without changing the standard one. CogPO is an OWL file: it can be reasoned over, but nothing was built to annotate data with it, and its authors’ hope of annotating experiment scripts with CogPO terms was not realized.
What CogPO has that HED does not¶
The paradigm layer. CogPO names 83 paradigms and says, for each, which condition pattern makes an experiment an instance. HED has none: it dropped its Paradigm hierarchy in the third generation and kept only the identifier slot, so a HED dataset says what happened but not which standard task it was. The Catalog exists to supply that layer for HED, and its task mapping to CogPO is where the two paradigm lists are compared.
A few items and presentations. HED 8.4.0 has no interoceptive presentation, no item for a digit or number, for food, an odor, a puff of air or speech as a stimulus, and no tap, draw, drink or name action. CogPO has all of these, because BrainMap experiments used them.
“In at least one of the conditions.” CogPO’s definitions are written for coding a whole experiment from a paper, so a class applies when any condition fits and overlapping classes are co-coded. HED has no equivalent coarse-grained coding of an experiment, by design: its unit is the event.
Gaps worth closing in the HED schema¶
Tagging CogPO’s vocabulary found the following absent from HED 8.4.0. Each is a candidate proposal to hed-schemas; whether to raise them is a decision for the HED Working Group.
A presentation tag for interoception beside
Somatic-presentationandVestibular-presentation.Items: a digit or number, food, an odorant, speech as a stimulus distinct from the
Speakaction, a checkerboard, a fixation point.Actions: tap, draw, drink, name, choose or decide, rate.
Body parts:
ArmandLegas parents of the existing upper-arm, forearm, upper-leg and lower-leg tags.
The three resources together¶
Resource |
Unit |
Says |
|---|---|---|
The Catalog |
a task |
what the participant was asked to do, and which processes that engages |
CogPO, and the Catalog’s facets |
a condition |
what was shown, in which modality and role; what was done, with which body part; under what instruction |
HED |
an event |
what happened, when, to whom, in which role |
Turner and Laird kept cognitive processes out of CogPO on purpose, because “if these links are explicit within the ontology then the ontology must be reconstructed or refactored every time the research community consensus about cognitive processes changes”, and left them to the Cognitive Atlas. The Catalog takes the other side of that bargain: it links every task to the processes it engages, keeps the links revisable, and grounds task identity in procedure, which is the part CogPO showed does not shift.
A dataset tagged with a Catalog task, whose task carries facet values, whose events are
annotated in HED, is described at all three levels with one vocabulary running through
them: the facet value words is the HED tag Word, the facet value button_press is
Press, and the task that has them is the one CogPO calls Stroop Task Paradigm.