What Is a Concept?
Published:
We encounter individual things.
This dog.
That chair.
One storm.
Yet we think in categories:
- dog,
- chair,
- storm.
A concept is part of what lets many different experiences count as instances of one kind.
Without concepts, general thought would be difficult.
Categorization
To categorize is to treat different things as equivalent for some purpose.
A Labrador and a poodle look different.
Yet both are recognized as:
dog.
Concepts organize variation.
Concepts Compress Experience
Instead of storing every object as unrelated, we group them.
This provides compression.
From:
millions of individual encounters,
we form:
general expectations.
Concept as Definition
A traditional view treats concepts as definitions.
A category is determined by necessary and sufficient conditions.
Example:
triangle = closed plane figure with three straight sides.
This works well for some formal categories.
Necessary Condition
A necessary condition must hold for every member.
Having three sides is necessary for being a triangle.
Sufficient Condition
A sufficient set of conditions guarantees category membership.
For Euclidean plane figures:
closed + three straight sides
is sufficient for triangle.
Natural Categories Are Harder
What are the necessary and sufficient conditions for:
game?
bird?
chair?
Ordinary concepts often resist sharp definitions.
Wittgenstein and Family Resemblance
Ludwig Wittgenstein famously discussed the concept:
game.
Games may share overlapping similarities without one feature common to every case.
He called this family resemblance.
Categories can have fuzzy structure.
Prototype Theory
Prototype theory suggests concepts are organized around typical examples.
A robin may be judged a more typical bird than a penguin.
Membership and typicality separate.
Prototype
A prototype is an abstract or representative center of a category.
New objects are compared with it.
The prototype need not correspond to one real object.
Typicality Effects
People respond faster to typical examples.
For example:
“robin is a bird”
may be verified faster than:
“ostrich is a bird.”
This suggests cognitive categories are not flat lists.
Exemplar Theory
Another view says categories are represented through remembered examples.
To classify a new object, compare it with stored exemplars.
Concept knowledge may be distributed over experience.
Prototype vs Exemplars
These theories are not necessarily exclusive.
Humans may use:
- prototypes,
- exemplars
depending on domain and task.
Cognition can combine strategies.
Theory-Theory
The theory-theory view says concepts are embedded in causal knowledge.
A concept like:
bird
includes expectations about:
- biology,
- behavior,
- development.
Concepts are mini-theories, not just similarity clusters.
Why Causal Knowledge Matters
A painted dog does not become a tiger because it looks striped.
We classify using beliefs about underlying nature.
Appearance alone is insufficient.
Essentialism
Humans often assume categories have hidden essences.
A tiger remains a tiger after shaving its fur.
This psychological essentialism can be useful.
It can also mislead when categories lack real essences.
Natural Kinds
Some categories may track objective structures in nature.
Examples include:
- chemical elements,
- biological taxa
under appropriate scientific definitions.
These are often called natural kinds.
Social Categories
Other categories are shaped partly by:
- institutions,
- norms,
- history.
Examples include many legal and social classifications.
Concepts do not all have the same metaphysical status.
Concept Boundaries
Some categories are sharp.
An integer either is or is not even.
Others have borderline cases.
Is a tomato a vegetable?
The answer depends on:
- botanical,
- culinary,
- legal context.
Context Dependence
Concept application changes with purpose.
A “small” elephant is large compared with a dog.
Concepts are interpreted relative to comparison classes.
Concepts and Language
Words express concepts.
But word and concept are not identical.
Different languages divide conceptual space differently.
One concept may have:
- multiple words,
- no exact translation.
Can Thought Exist Without Words?
Yes, likely.
Infants and animals show categorization without full language.
Language can refine and stabilize concepts.
It is not their only possible medium.
Concepts and Perception
Perception is shaped by learned categories.
An expert radiologist sees distinctions a novice misses.
Concepts influence what becomes salient.
Categorical Perception
Continuous differences can be perceived as discrete categories.
Speech sounds provide classic examples.
Conceptual boundaries can reshape perceptual judgment.
Concepts as Representations
In symbolic theories, a concept may be represented by a symbol or structured description.
In neural theories, a concept may be distributed over activation patterns.
The implementation question remains open.
Semantic Networks
Concepts can be represented in networks.
Nodes represent concepts.
Edges represent relations such as:
- is-a,
- part-of,
- causes.
This captures relational knowledge.
Frames
A frame organizes stereotypical knowledge.
Concept:
restaurant
may include roles such as:
- customer,
- server,
- menu,
- bill.
Concepts contain structured expectations.
Scripts
A script represents typical event sequences.
Restaurant script:
- enter,
- order,
- eat,
- pay.
Scripts help interpret incomplete situations.
Conceptual Dependency
Some AI systems attempted to represent meaning using primitive conceptual relations.
The goal was to move beyond surface words.
These symbolic approaches influenced knowledge representation.
Neural Concepts
Neuroscience suggests concepts are distributed across multiple systems.
Thinking about:
hammer
may recruit representations related to:
- shape,
- action,
- use.
Concepts can be grounded in sensorimotor networks.
Hub-and-Spoke Models
Some semantic-memory theories propose:
- modality-specific “spokes,”
- integrative “hub” regions.
This architecture aims to explain how distributed features become coherent concepts.
Concepts as High-Dimensional Geometry
Modern machine learning often represents concepts as vectors.
Related items occupy nearby regions in a learned space.
This gives concepts geometric structure.
Embeddings
An embedding might place:
king, queen, man, woman
in a vector space with systematic relationships.
Such spaces capture statistical structure.
They do not automatically reproduce human conceptual understanding.
Concept Drift
Concepts can change over time.
A classifier trained on yesterday’s data may fail when category boundaries shift.
This is called concept drift in machine learning.
Concepts are not always fixed.
Scientific Concepts
Science advances partly by inventing new concepts.
Examples include:
- gene,
- field,
- entropy,
- spacetime.
A new concept can reorganize many observations at once.
Conceptual Revolution
Sometimes progress requires replacing old categories.
“Phlogiston” disappeared.
“Heat” changed from substance-like fluid to statistical molecular motion.
Concepts shape theory.
Bad Concepts
A concept can mislead if it groups things that should be separated.
Scientific progress may require abandoning a familiar category.
Conceptual engineering is part of knowledge.
Concepts and Prediction
A useful concept supports expectations.
If something is a bird, we predict:
- biological properties,
- behavior.
Concepts are compressed models.
Concepts and Action
Categories also guide action.
If an object is categorized as:
food,
we behave differently than if it is categorized as:
tool.
Concepts connect perception to goals.
Concepts and Identity
Self-concepts influence behavior.
A person may think:
“I am a scientist.”
That category changes:
- expectations,
- goals,
- social interpretation.
Concepts can organize the self.
Concepts and Bias
Human categories can generate stereotypes.
When useful statistical summaries are overgeneralized to individuals, reasoning becomes distorted.
Categorization is necessary.
Its misuse is dangerous.
Machine Categories
Machine-learning systems classify inputs using learned boundaries.
These categories may not align with human concepts.
A model can distinguish patterns without representing the same conceptual structure humans do.
Adversarial Examples
Tiny input changes can cause some classifiers to change category while humans see no meaningful difference.
This shows that machine representations can carve space differently from human concepts.
Grounding
What gives a concept meaning?
Possibilities include:
- relations to other concepts,
- sensory experience,
- action,
- causal interaction.
The symbol-grounding problem appears again.
Concepts Are Tools
Concepts are not passive labels.
They help us:
- compress,
- predict,
- infer,
- communicate.
Their value lies partly in what they allow us to do.
The Philosophical Lesson
A concept is not always a strict definition.
Concepts may involve:
- prototypes,
- exemplars,
- causal theories,
- relational structures.
They organize experience by treating different things as relevantly similar.
The Next Question
If concepts and thoughts are represented in the brain, how?
Does one neuron represent one idea?
Do concepts correspond to patterns across populations?
What does it even mean for neural activity to represent something?
The next essay asks:
Neural Activity and Mental Representation.