What is in the Cognitive Atlas

The Cognitive Atlas is a collaboratively built ontology of cognitive processes and the experimental paradigms that measure them. It separates concepts (mental processes) from tasks (experimental paradigms), gives each a stable identifier, and records which concepts a task is claimed to assess. It is the only openly licensed lexicon spanning the breadth of cognitive neuroscience, and it was the starting corpus for this catalog.

This page describes the Atlas on its own terms: what it contains, how completely its entries are filled in, and what condition the data is in. It draws no comparison with the task and process catalog published here.

In short

The Atlas is best understood as a broad but thinly and unevenly curated corpus rather than a finished reference. Its breadth is real, and so is its architecture: separating processes from paradigms is right, the identifiers are stable, and a genuine 1398-edge relation graph connects concepts to one another.

The curation is partial, in both layers at once. On the task side, 298 of 857 tasks (34.8%) assert no concept at all, which disconnects them from the ontology that gives the Atlas its value, and only 363 (42.4%) have a definition, a concept, and a citation together. On the concept side, 455 of 918 concepts (49.6%) are asserted by no task, 413 (45.0%) carry no class, and 774 (84.3%) have no citation. Curation also stopped: almost nothing in either layer was entered after 2017.

Anything built on the Atlas should treat it as a source of candidate terms and stable identifiers to be verified, not as a curated authority.

What was captured

The figures come from a snapshot of the Atlas REST API (http://cognitiveatlas.org/api/v-alpha) taken on 2026-09-16. Both layers were pulled in full: the bulk listing for each, then the detail record for every entry, which is where concept relations and citations live.

Layer	Entries	Detail records retrieved
Tasks (experimental paradigms)	857	857
Concepts (mental processes)	918	918

Pulling the concept endpoint directly matters. A harvest taken from the task endpoint alone reaches a concept only when some task asserts it, which hides 455 of the 918 concepts and every one of the concept-to-concept relations.

Tasks

Definitions

Definition length	Tasks	Share
Missing or placeholder	38	4.4%
Under 50 characters	21	2.5%
50 to 199	316	36.9%
200 to 499	294	34.3%
500 to 999	144	16.8%
1000 or more	44	5.1%

Among the 819 tasks that have a real definition, the mean length is 362 characters, the median 248, and the longest 4074. The spread is the point: entries range from a single clause to a small essay, with no house style.

36 of the entries counted as missing hold the literal four-character string None rather than any text. They cover well-known paradigms, among them visual search task, serial reaction time task, and color naming task.

Concept linkage

Concepts asserted	Tasks	Share
0 concepts	298	34.8%
1 concept	249	29.1%
2 to 3	189	22.1%
4 to 7	93	10.9%
8 or more	28	3.3%

There are 1422 task-to-concept links in total, a mean of 1.66 per task against a maximum of 16. About a third of tasks (34.8%) assert no concept whatsoever and another 29.1% assert exactly one, so for roughly 64% of entries the knowledge graph is either absent or a single edge. A long definition is no guarantee of a linked one: several entries carry a substantial write-up and no concepts at all.

Citations

Measure	Count	Share of tasks
Tasks with at least one citation	470	54.8%
Tasks with no citation	387	45.2%

The 470 cited tasks carry 1091 citations between them, of which 946 record a PubMed id. The rest are reachable only through a free-text reference or a bare URL.

Everything filled in at once

Combination	Tasks	Share
Definition and concepts and citation	363	42.4%
No citation and no concepts	203	23.7%
No definition and no concepts	12	1.4%

Fewer than half of all task entries are complete in this minimal sense.

Other structured fields

The schema offers more than definitions, concepts, and citations. Most of it is sparsely used.

Field	Tasks with at least one	Share	Total entries
Contrasts	529	61.7%	1585
Indicators	317	37.0%	584
Conditions	272	31.7%	807
Batteries	113	13.2%	134
External datasets	66	7.7%	81
Implementations	55	6.4%	66
Disorders	16	1.9%	20

Contrasts are the exception and are populated for most tasks, which makes them the most reusable structured content in the Atlas after the concept links themselves.

Concepts

Definitions

Definition length	Concepts	Share
Missing or placeholder	23	2.5%
Under 30 characters	19	2.1%
30 to 149	602	65.6%
150 or more	274	29.8%

Concept definitions are shorter and more consistently present than task definitions. 23 are unfinished in a way that is visible in the published data, carrying an editing placeholder such as ADD DEFINITION HERE, or a serialized null, in place of a definition:

Active Cognitive Inhibition, Limited Capacity, arousal, behavioral inhibition, decision uncertainty, emotional reappraisal, emotional self-evaluation, exogenous attention, face maintenance, goal selection, implicit learning, insomnia, localization, negative emotion, numerical scale judgment, overt naming, phonological assembly, phonological comparison, resistance to distractor inference, risk aversion, social inference, test term, working memory updating.

These are not obscure corners of the vocabulary. working memory updating, implicit learning, arousal, risk aversion, and exogenous attention are all constructs in active use, published with no definition at all.

The half no task points at

Group	Concepts	Share	No definition	No class	Median definition
Asserted by at least one task	463	50.4%	2.4%	38.9%	102 chars
Asserted by no task	455	49.6%	2.6%	51.2%	102 chars

Nearly half the concept layer is asserted by no task at all. The instinct is to assume those are the leftovers, but they are not: the two groups have the same median definition length and nearly the same rate of missing definitions. The orphans are ordinary, adequately defined concepts that simply never got wired to a paradigm. They are somewhat less likely to carry a class, but the difference is one of degree.

The consequence is that the Atlas’s task-to-concept graph rests on about half its own vocabulary, and a consumer who reaches the Atlas through tasks never sees the rest.

Concept classes

The Atlas sorts concepts into ten top-level classes. Class assignment is the single largest gap in the concept layer: 413 of 918 concepts (45.0%) carry no class.

Concept class	Concepts	Share
(no class assigned)	413	45.0%
Learning and Memory	116	12.6%
Language	94	10.2%
Perception	87	9.5%
Reasoning and Decision Making	64	7.0%
Executive/Cognitive Control	42	4.6%
Attention	37	4.0%
Emotion	33	3.6%
Social Function	13	1.4%
Action	10	1.1%
Motivation	9	1.0%

This is not confined to rarely used terms. The most heavily used concepts carrying no class are:

Concept	Tasks asserting it
impulsivity	17
Limited Capacity	11
emotion regulation	10
restricted behavior	10
social motivation	9
cognitive development	6
defiance	6
hyperactivity	6
processing capacity	6
future time	5
interference control	5
obsession	5

Among the classified concepts the balance is skewed. Learning and Memory, Language, and Perception dominate, while Motivation, Action, and Social Function are barely populated. That reflects curator interest rather than the shape of the field.

The class layer also mixes kinds of thing. Alongside cognitive processes it carries traits, symptoms, and clinical constructs such as impulsivity, hyperactivity, defiance, obsession, anhedonia, perfectionism, and restricted behavior. These are legitimate research constructs but they are not mental processes, and nothing in the schema separates them from those that are.

Relations between concepts

Unlike the flat class assignment, the concept layer carries a real relation graph: 1398 edges over 562 concepts.

Relation	Edges
KINDOF	866
PARTOF	532

This is the Atlas at its most valuable and is invisible to anyone who reads only the task endpoint. It is also incomplete: 356 concepts (38.8%) sit in the graph with no relation to any other concept, so the hierarchy covers a majority of the vocabulary but far from all of it.

Citations

Only 144 concepts (15.7%) carry any citation, together holding 202 references. The other 774 concepts (84.3%) have a definition with no source of record, which is the concept layer’s most consequential omission for anyone who needs to justify a term.

How widely concepts are used

Tasks per concept	Concepts	Share
0 tasks	455	49.6%
1 task	234	25.5%
2 tasks	87	9.5%
3 to 9 tasks	118	12.9%
10 or more tasks	24	2.6%

A small head does most of the work:

Concept	Tasks
visual perception	40
attention	29
response selection	28
working memory	27
cognitive control	27
motor control	26
response execution	24
auditory perception	21
language	20
response inhibition	19
impulsivity	17
active maintenance	15
spatial ability	13
language comprehension	13
attentional focusing	12

These are the coarsest available labels. visual perception, attention, and cognitive control are the terms an annotator reaches for when nothing more specific is at hand. Their dominance suggests annotation stopped at the top of the hierarchy rather than showing that these processes matter most.

Overall task annotation quality

Combining definition length with concept count sorts every task entry into a tier. An entry is well annotated with a definition of 300 characters or more and at least four concepts, and skeletal with a definition under 50 characters or no concepts at all.

Tier	Tasks	Share
Well annotated	54	6.3%
Adequate	410	47.8%
Minimal	63	7.4%
Skeletal	330	38.5%

Only 54 entries (6.3%) are well annotated, while 330 (38.5%) are skeletal.

What kind of thing is an entry?

The Atlas files everything under “task”. In practice the corpus mixes experimental paradigms with instruments that are not paradigms at all. Classifying by name gives a lower bound, since it only catches entries whose name declares what they are.

Entry kind (name-based)	Entries
Rating scale, questionnaire, or inventory	107
Standardized test or battery	45
Imaging protocol or localizer label	29
Stimulation or physiological procedure	19
Not matched (largely experimental paradigms)	657

At least 200 entries are rating scales, questionnaires, standardized batteries, imaging protocol labels, or physiological procedures. Nothing in the record distinguishes them from experimental paradigms, so any consumer has to impose that distinction itself.

Duplicate and fragmented entries

A paradigm family is often spread across several entries, one per implementation, with no entry marked canonical and the best-annotated one frequently not the standard version.

Family	Entries	With zero concepts	Best-annotated entry	Concepts
naming	16	4	stop signal task with letter naming	5
span	13	1	backward digit span task	4
stroop	8	2	Stroop task	4
n-back	8	2	n-back task	8
stop signal	8	1	conditional stop signal task	9
fluency	6	4	category fluency test	4
sternberg	4	1	Sternberg Recent Probes	4
continuous performance	4	2	Penn continuous performance task	3
oddball	3	2	oddball task	6
weather prediction	3	1	dual-task weather prediction	13
bandit	3	3	Volatile Bandit	0

A reader searching for a paradigm lands on whichever variant matches their wording, and the quality of what they find is largely accidental.

Names are duplicated outright in both layers: boston naming test, false belief task among tasks, and autobiographical memory, cognitive warfare, implicit learning, risk aversion among concepts. One concept is named test term, with no definition, and is published alongside the rest.

Data hygiene

Issue	Tasks	Concepts
Raw HTML entities in the definition	93	59
Mis-decoded UTF-8 in the definition	87	22
Definition is the literal string None	36	6

Definitions were pasted in from mixed sources without normalization, so escaped markup such as ' and " survives in the published text, along with byte sequences from a double-encoding error. Any text taken from the Atlas needs cleaning before it is displayed.

Non-ASCII characters in entry names are worth special care. Several tasks use a curly apostrophe or an en dash in their name (Raven's Progressive Matrices Test, Penn's Logical Reasoning Test, Angling Risk Task - Always Sunny), which is a common source of retrieval failures in client code that assumes ASCII.

When the curation happened

Year	Task entries	Concept entries
(none)	115	103
2009	101	538
2010	85	9
2011	97	22
2012	117	51
2013	105	58
2014	5	1
2015	143	122
2016	70	2
2017	19	12

Entry timestamps cluster in an initial build-out around 2009 and a second push in 2012 and 2015, then stop. The absence of paradigms that became standard afterwards is a direct consequence, and so is the absence of the computational vocabulary (model-based and model-free learning, reward prediction error, evidence accumulation) that the field adopted over the same period.

Reading the Atlas fairly

The weaknesses above are those of an unfunded community resource that stopped being actively curated, not of its design. What the Atlas got right still matters: separating processes from paradigms is the correct architecture, the identifiers are stable and citable, the relation graph is real, and the breadth of coverage is unmatched by any open alternative. For the subset of entries that were curated properly, the task-to-concept graph is exactly the structure a paradigm ontology needs.

The practical conclusion is about how to use it. The Atlas is a well-designed, broadly scoped, partially populated corpus. It is an excellent source of candidate paradigm names, concept labels, and stable identifiers. It is not a source that can be consumed without verification, because a given entry may be complete, a stub, a duplicate variant, or a questionnaire, and nothing in the record says which.

Reference

Poldrack, R. A., Kittur, A., Kalar, D., Miller, E., Seppa, C., Gil, Y., Parker, D. S., Sabb, F. W., & Bilder, R. M. (2011). The Cognitive Atlas: Toward a knowledge foundation for cognitive neuroscience. Frontiers in Neuroinformatics, 5, 17.

The Atlas is published at https://www.cognitiveatlas.org/, with a REST API at http://cognitiveatlas.org/api/v-alpha and a Python client at https://github.com/CognitiveAtlas/cogat-python.