Generalisation and Discrimination

Learning would not be very useful if it worked only once.

Imagine learning that one particular cue predicts something important.

Then the cue changes slightly.

Its colour is different.

Its shape changes.

It appears somewhere new.

If prior learning influenced behaviour only under conditions closely matching the original experience, its usefulness under natural variation would be severely limited.

But the opposite would create another problem.

If anything vaguely similar produced the same response, important differences would disappear.

A situation that looks familiar may predict a completely different outcome.

A small distinction may be exactly what matters.

Learning therefore faces a problem beyond simply acquiring a relationship.

Learning is acquired under conditions that already help shape where its influence extends, and later experience can continue to reshape that scope.

Where does this learning apply?

Two complementary processes help explain that scope: generalisation, and discrimination.

One gives learning reach.

The other gives it selectivity.

Neither is inherently better.

Generalisation Gives Learning Reach

As established in D5.2 — Associative Learning, experience with relationships among events can alter subsequent behaviour.

But the behavioural influence of that learning does not have to remain confined to the exact conditions under which it was acquired.

Suppose an organism learns that a particular stimulus predicts an outcome.

Later, it encounters a stimulus that differs from the original in some way.

If the previous learning influences responding to this new stimulus, the learning has generalised.

In broad terms:

Generalisation occurs when learning acquired under one set of conditions influences responding under other conditions.

Stimulus generalisation is one familiar example, and D5.3 — Classical Conditioning supplies its clearest architecture.

A response learned in relation to one stimulus can extend to other stimuli sharing relevant properties with it.

This is not automatically a failure of precision.

generalisation ≠ error

Without some capacity for generalisation, learning would have very little reach.

The world rarely repeats itself perfectly.

Objects appear from different angles.

Situations contain different details.

Conditions change.

New examples appear.

For prior experience to remain useful, learning must sometimes influence behaviour despite those differences.

Generalisation is one way it does.

But generalisation is not the complete theory of transfer.

Generalisation is one way prior learning can influence new conditions; broader transfer of skills, strategies or abstract knowledge can involve additional processes.

But Similar Does Not Always Mean Equivalent

The same capacity creates another problem.

Two situations can look similar while predicting different outcomes or consequences.

Imagine that an organism responds similarly to two cues.

Then experience establishes that one continues to predict an important outcome while the other does not.

The physical difference between the cues may have existed all along.

But experience has now made that difference predictive and behaviourally consequential.

Responding can become increasingly selective.

This is discrimination learning.

Discrimination is not merely the ability to perceive that two things differ.

An organism may be capable of detecting a difference without having learned that the distinction matters for the response in question.

Instead:

Discrimination develops when experience makes differences among stimuli, situations or contexts behaviourally consequential.

Consequences can be especially important here.

As established in D5.4 — Operant Learning, behaviour changes partly through its consequences. If apparently similar situations are associated with different consequences, experience can make their differences increasingly important for action.

But discrimination does not require someone to consciously reason:

"I have identified the relevant distinction."

Learned responding can become selective without that distinction being explicitly formulated.

discrimination learning ≠ conscious reasoning

Generalisation and Discrimination Solve the Same Problem

Generalisation and discrimination can sound like opposites.

One broadens responding.

The other makes it more selective.

But that framing misses their shared function.

Both help answer: which differences should matter for behaviour, and which should not?

More precisely: which differences predict relevantly different outcomes or consequences, and which variations can be treated alike for the behaviour in question?

If a difference does not alter anything relevant, treating two situations similarly can be useful.

If a difference predicts a different outcome or consequence, treating them identically can be costly.

Generalisation allows learned responding to extend across variation.

Discrimination allows that responding to become selective when variation matters.

generalisation + discrimination = complementary scope-setting processes

Here, scope-setting is explanatory language.

It does not imply that two separate mechanisms consciously negotiate or calculate a boundary.

It describes the pattern of conditions across which learning influences behaviour.

This is why discrimination is not automatically a more advanced form of learning.

And generalisation is not automatically a less precise one.

The useful scope depends on the environment.

A learning system that discriminated every tiny difference would transfer poorly.

A learning system that generalised across every resemblance would miss consequential distinctions.

The adaptive problem is not to maximize either process.

It is to develop a scope of responding that fits the relevant structure and consequences of the environment.

That fit will not always be perfect.

Learning occurs through limited experience in environments that can be noisy and change over time.

But the problem remains the same: which variations can be treated alike, and which predict something different enough to change behaviour?

Similarity Matters—but It Is Not the Whole Explanation

Similarity clearly matters to generalisation.

If learning occurs in relation to one stimulus, responding often extends more strongly to some stimuli than to others.

But saying "learning generalises to similar things" leaves an important question unanswered.

Similar in what way?

Physical resemblance can matter.

Perceptual organization can matter.

Contextual relationships can matter.

And learning history can make conditions that look different support similar responding—or make conditions that look similar support different responding.

These are not simply four interchangeable similarity scales.

They are different ways in which the relationship between current conditions and previous learning can influence behaviour.

Two stimuli may look almost identical but repeatedly predict different outcomes.

Learning can make their difference highly consequential.

Two situations may look quite different but repeatedly support the same relevant relationship or consequence.

They can come to support similar responding.

This is why:

perceptual similarity ≠ functional equivalence

The broader principle in D6.1 — Perception Is Constructive is useful here.

What becomes behaviourally relevant cannot always be read directly from physical resemblance alone.

But this does not mean similarity is arbitrary.

Physical and perceptual properties can strongly constrain generalisation.

The point is that similarity is part of the explanation, not the entire explanation.

Learning history matters too.

Experience Can Change the Boundary

Suppose two similar stimuli initially produce similar responding.

Now experience repeatedly establishes that they predict different outcomes.

Over time, behaviour may become more differentiated between them.

The stimuli did not suddenly become physically different.

What changed was the behavioural significance that experience had given to the difference between them.

The difference now predicts something relevant to responding.

This demonstrates something important:

physical difference ≠ automatically behaviourally relevant difference

Experience can make a distinction matter.

And once it matters, the pattern of generalisation can change.

Discrimination experience can reshape which stimuli produce similar responding and which produce different responding.

This does not mean discrimination always makes generalisation narrower.

Different procedures can produce different patterns.

The broader point is that experience can reshape the conditions across which learned responding occurs.

That is what it means to describe learning as having an experience-dependent boundary or scope.

The boundary is not a literal line stored somewhere inside the organism.

Nor does it necessarily appear only after acquisition.

Learning occurs under conditions that already shape its scope, and subsequent experience can continue to reshape the pattern of conditions across which it influences behaviour.

Generalisation Gradients Show Behavioural Scope

Researchers can make aspects of this scope visible experimentally.

Suppose learning occurs with one particular stimulus, as in the conditioning procedures of D5.3 — Classical Conditioning.

Later, several related stimuli are presented, varying along some dimension.

Researchers can measure how strongly the organism responds to each.

The resulting pattern is called a generalisation gradient.

For example, a gradient may show strong responding near the trained stimulus and systematically different responding as the test stimuli vary.

Other patterns are possible.

The gradient therefore needs to be interpreted carefully.

It tells us how measured behaviour varied across the tested conditions.

It does not directly show the complete structure of what the organism represents internally.

generalisation gradient ≠ transparent map of internal representation

The shape of the gradient can depend on more than physical similarity.

Previous discrimination experience can change it.

The dimensions emphasized during learning can matter.

Context can matter.

Testing conditions can matter.

So a generalisation gradient is valuable evidence about behavioural scope.

It is not a direct picture of the learning itself.

Broad Is Not Strong and Narrow Is Not Better

This distinction matters because it is easy to treat the width of generalisation as a measure of learning quality.

Suppose responding generalises broadly.

Does that mean the learning was stronger?

Not necessarily.

If all those situations share the same relevant structure or predict the same relevant outcome, broad generalisation may be exactly what is useful.

But if some predict importantly different outcomes or consequences, the same breadth may prevent useful discrimination.

Now suppose responding is very narrow.

Does that mean the learning is more precise?

Again, not necessarily.

Narrow responding can be useful when small differences predict relevantly different outcomes.

But it can also prevent learning from extending across variation that is irrelevant to the task or behaviour.

So:

breadth of generalisation ≠ learning strength

And:

narrowness of generalisation ≠ learning quality

Terms such as overgeneralisation and undergeneralisation only make sense relative to the relevant task, outcomes and environmental structure.

Broad responding becomes overgeneralisation when it extends across conditions that predict relevantly different outcomes or consequences and would therefore be useful to distinguish for the behaviour in question.

Narrow responding becomes undergeneralisation when learning fails to extend across conditions whose differences are irrelevant to the relevant task or consequence structure.

The width alone does not tell us which one we are looking at.

adaptive value depends on fit

Context Helps Define Where Learning Applies

The scope of learning can also extend across—or remain sensitive to—context.

A response acquired under one set of conditions may not be expressed identically under another.

This does not automatically mean that the original learning disappeared or was weak.

It means that acquisition and generalisation are different questions.

As established in D5.1 — What Is Learning?, what has been learned and what is expressed in behaviour at a particular moment should not be collapsed.

failure to transfer ≠ absence of learning

But failure to transfer is still informative.

It can provide evidence about the limits of generalisation or behavioural expression under the tested conditions without, by itself, proving that acquisition never occurred.

Likewise:

observed transfer ≠ complete measure of learning

If behaviour generalises broadly across several contexts, that tells us where prior learning influenced behaviour under those conditions.

It does not automatically tell us that the learning was stronger, deeper or more abstract.

The adjacent architecture in D5.6 — Extinction Is New Learning makes this especially visible.

Extinction learning can genuinely occur under one set of conditions without controlling responding identically everywhere else.

The important point for this piece is not the extinction mechanism itself.

It is the scope problem: learning can occur without generalising completely across conditions.

Context is therefore not merely a background variable added after learning has already been retrieved unchanged.

It can participate in the conditions under which learning is acquired and expressed.

And identifying those conditions still does not explain the whole behaviour.

As D1.8 — Causes, Conditions, Triggers and Constraints makes clear, identifying the conditions across which learned responding generalises does not by itself provide the complete causal explanation of behaviour.

scope of learned responding ≠ complete causal architecture

The Grounded Insight

Learning does more than establish a response.

It develops under conditions that help shape where its behavioural influence extends.

That scope can continue to change through later experience.

A learned relationship can influence behaviour beyond the exact circumstances in which it was acquired.

That gives learning reach.

But useful behaviour cannot treat every resemblance as equivalence.

Experience can make differences matter.

That gives learning selectivity.

Similarity matters.

So do outcomes and consequences.

So does learning history.

So does context.

And none of these implies that the organism always discovers the perfect boundary.

The adaptive problem is to develop a scope in which learned responding extends across differences that do not alter relevant outcomes or consequences while remaining selective to differences that do.

That is why:

generalisation ≠ error

And:

discrimination ≠ inherently better learning

They are complementary parts of the same adaptive problem.

A useful conceptual map is this. Learning history — relationships, consequences and prior discrimination experience — meets the current stimulus, situation and context, the conditions under which behaviour is now expressed. Together they produce the behavioural scope under current conditions: where prior learning influences responding, and where responding differentiates. Generalisation is the side of that scope where learned influence extends across variation. Discrimination is the side where learned responding becomes selective because differences matter.

This is not a literal decision tree inside the organism.

Nor does it imply that scope is first created only when a new situation appears.

It describes how learning history and current conditions jointly constrain where learned influence becomes behaviourally evident.

learning includes learning where a response applies

That scope is not fixed by resemblance alone.

It can be reshaped by experience.

And whether it is useful depends not on how broad or narrow it is, but on how well the distinctions it preserves—or treats alike—fit the relevant structure and consequences of the environment.