What is a model?
A model is a simplified representation of something.
A map models a place. A globe models the Earth. A weather forecast models how the atmosphere may change. A belief about another person's character is also a model, though we may not call it one.
A model preserves selected relationships while leaving other details out.
It is useful not because it contains everything, but because it contains enough for a purpose.
Why must a model leave things out?
Reality contains more detail than any mind or machine can use at once.
A street map shows roads and names but usually omits the shape of every leaf. That omission is not a defect when the purpose is navigation.
If the purpose changes, the useful details change. A drainage engineer and a tourist need different maps of the same city.
Every model therefore answers an often hidden question: useful for what?
Is a model a copy of reality?
No.
A copy attempts to reproduce. A model selects, organizes and interprets.
A toy aircraft may resemble a real aircraft but fail to explain how lift works. A few equations may look nothing like an aircraft yet predict its motion well.
Resemblance is one way to represent, but structure and relationship often matter more.
Confusing a model with the thing itself makes its omissions invisible.
What kinds of model are there?
Models take many forms:
- Physical models — a globe, prototype or scale building
- Conceptual models — an explanation of how parts relate
- Mathematical models — quantities and equations describing a pattern
- Statistical models — relationships estimated from observations
- Mental models — the internal expectations used in thought and action
- Computational models — representations whose behaviour is simulated by a computer
The forms differ, but each stands in for something so that it can be examined more easily.
How does a model help us predict?
A model connects present information with possible future states.
Dark clouds, falling air pressure and earlier experience may support a model in which rain is likely. The prediction does not come from the clouds alone. It comes from a learnt relationship between observations and outcomes.
Prediction is a test of the model.
When the expected rain does not arrive, the failure provides information. Perhaps the model ignored wind, location or season. A useful model can be corrected by the gap between expectation and reality.
How does a model explain?
An explanation identifies relationships that make an event less surprising.
Saying that a lamp failed because “it stopped working” adds little. A model of the circuit suggests possible causes: no electrical supply, a broken wire, an open switch or a damaged bulb.
The model explains by showing how changes in one part affect another.
A strong explanation also supports new predictions. If the bulb is the cause, replacing it should restore the light.
How are models learnt?
A system compares experiences and preserves recurring patterns.
A child learns that unsupported objects fall. A mechanic learns how a particular sound relates to a failing bearing. A machine-learning system adjusts internal values so that its outputs better match examples.
Learning is not merely adding facts. It is changing a model so that future interpretation and action improve.
The learner must decide which similarities matter. That is why more experience does not always produce a better model.
Can a model be wrong and useful?
Yes.
A flat street map is wrong if treated as a complete description of the curved Earth. It is still useful for walking across a neighbourhood.
Many models are approximations. They work within a limited range and fail outside it.
The question is not simply “Is this model true?” but:
- Which parts of reality does it preserve?
- How accurate is it?
- Under what conditions does it work?
- What decision will depend upon it?
A model may be good enough for one purpose and dangerously poor for another.
What is overfitting?
Overfitting happens when a model follows the accidents of its examples instead of the wider pattern.
Imagine memorizing every answer in one practice paper without understanding the subject. Performance on that paper will be excellent, but a new question may cause failure.
A statistical model can do the same. It may match its training data closely while making poor predictions about new cases.
A model must learn from experience without becoming trapped inside the exact experiences it has seen.
What is underfitting?
Underfitting is the opposite failure.
The model is too simple to capture an important pattern. Predicting the same temperature every day ignores season, time and weather.
Simplicity is valuable because it makes a model easier to understand and apply. But excessive simplicity removes the very relationships needed for prediction.
Good modelling balances useful detail against unnecessary complexity.
How do models shape perception?
We do not first receive a complete world and then build a model of it.
Cognition uses earlier models while interpreting present signals. A rope in dim light may be seen as a snake because expectation helps organize an uncertain image.
This makes perception fast. It also permits error.
What we notice, remember and believe is influenced by the models already present. Correcting a model may therefore change not only our conclusion, but what appears obvious.
How do models guide decisions?
A decision compares possible actions through predicted consequences.
Should I carry an umbrella? The answer depends upon a model of the weather, the cost of carrying it and the cost of getting wet.
The chosen action then produces new information. If the forecast repeatedly fails, the model or the trust placed in it should change.
Action is therefore both the use of a model and a way of testing it.
Is a machine-learning model a program?
It is part of a program, but the two are not identical.
The program defines how data is received, how the model is trained and how its output is used. The model contains adjustable relationships learnt from data.
Two copies of the same training program can produce different models when trained on different examples.
The program provides the learning procedure. The model is what that procedure has learnt.
So what is a model, finally?
A model is a selective representation used to explain, predict or act.
It compresses reality into relationships that a mind or machine can manage. Its omissions make it usable, but also give it limits.
We never act with all of reality in our hands. We act through models.
Intelligence requires building them, testing them and remembering that even the best map is not the land.
