Associative Learning

A cue repeatedly appears before an event. Later, the cue changes behaviour.

It is tempting to explain the learning by saying that the two things became associated because they occurred together. That intuition captures something real, but it is not enough.

If the same event occurs just as often when the cue is absent, the cue may carry little additional information. If another cue already predicts the event, a new cue presented alongside it may acquire less influence than repeated pairing alone would suggest. And when behaviour eventually changes, the response itself does not necessarily reveal exactly what relation was learned.

Associative learning concerns what experience changes about relationships, not merely what events occur together.

For the purposes of this Library, associative learning refers broadly to experience-dependent learning about relationships among cues, events, actions and outcomes, such that those learned relations can influence later anticipation or behaviour.

This describes the phenomenon. It does not commit us to one universal theory of how those relations are represented or updated.

Relationships, not just proximity

One of the oldest intuitions about association is contiguity: events that occur close together in time or space become linked.

Temporal proximity can matter. But proximity alone does not specify the relation an organism can learn.

A second concept is contingency. In general, contingency concerns the statistical relationship among events, cues, actions and outcomes, although its precise operational meaning differs across learning paradigms.

Consider a cue that is repeatedly followed by an outcome.

If the outcome rarely occurs when the cue is absent, the cue provides useful information about what is likely to happen next.

Now imagine that the same outcome occurs just as often without the cue.

The cue and outcome may still have been paired repeatedly. Yet the cue now provides much less information.

Classic contingency experiments demonstrated this clearly: what happened outside the cue–outcome pairings affected what was learned about the cue. Counting how often two events occurred together was not enough.

Contiguity and contingency therefore answer different questions.

Contiguity: did the events occur near one another?

Contingency: what relationship among the events made one informative about another?

The point is not that temporal proximity is irrelevant. It is that proximity alone cannot provide a complete account of associative learning.

Prior learning can complicate matters further. In cue-competition phenomena such as blocking, previous learning about one predictor can reduce later learning about another even when the new cue is repeatedly paired with the outcome. These effects have been theoretically influential, although their magnitude and robustness depend on procedure.

The lesson needed here is narrow: what is acquired from an experience depends on more than the presence of a pairing.

What kinds of relationships can be learned?

Associative learning is broader than a stimulus acquiring one fixed response.

One major kind of relation concerns events.

Through experience, a cue can become informative about another event or outcome:

cue → outcome

The relevant change may affect later anticipation or responding when the cue appears. This does not require the cue to produce one invariant behaviour, nor does the word prediction imply that the learner consciously formulates a prediction.

This broad event–outcome architecture provides the foundation for Classical Conditioning.

Associative learning can also concern the learner's own actions.

Experience can establish relationships between an action and a subsequent outcome:

action → outcome

Different actions can become related to different consequences, and those relations can later influence behaviour.

That does not yet tell us whether a consequence functions as reinforcement, punishment or reward. Nor does action–outcome learning itself define operant behaviour.

Those distinctions belong downstream in Operant Learning and Reinforcement, Reward and Pleasure.

The important point at D5.2's level is that associative learning can involve both relations among events and relations between actions and outcomes.

A learned predictive relationship should also not automatically be interpreted as a learned claim about causation. A cue may predict an outcome without causing it, and behavioural sensitivity to an action–outcome relation does not by itself tell us what causal understanding, if any, the learner represents. Those questions require additional evidence.

The relation in experience, the relation learned, and the response are not the same thing

What Is Learning? established that learning and performance should not be treated as identical. Associative learning adds another layer to that distinction.

Suppose experience contains a relation such as:

cue → outcome

We can distinguish three things analytically.

There is the relation present in experience: what events, actions and outcomes actually occurred and how they were arranged.

There is the relation learned: the sensitivity or information acquired through that experience.

And there is the behavioural expression: how the acquired learning influences expectation, choice or responding under current conditions.

The relation present in experience, the relation learned, and the behaviour used to measure it are related but not identical.

This is an explanatory distinction, not a claim that associative learning literally proceeds through three discrete stages.

A contingency can exist in the environment without the learner necessarily becoming behaviourally sensitive to it. What is learned need not reproduce the structure of experience perfectly, because prior learning and current conditions can affect acquisition. And the same learned relation need not produce the same response under every circumstance.

Here, what was learned is scientific shorthand for the information or relational sensitivity supported by the evidence. It does not assume one particular internal representational format.

That distinction matters because behavioural change after associative experience tells us that learning occurred only imperfectly; it does not automatically tell us the exact form the learning took.

Behaviour does not tell us everything that was learned

Suppose a cue changes behaviour after training.

One explanation might be that the cue became directly associated with a response.

Another possibility is that the cue became related to a particular outcome, and information about that outcome influences the response.

These possibilities can sometimes produce similar outward behaviour.

Researchers therefore use additional manipulations to distinguish among them. For example, changing the value of an outcome after learning can selectively change behaviour connected to that outcome. Such findings show that behaviour can depend on information about specific outcomes rather than only on a fixed stimulus–response connection.

Action–outcome research provides a parallel example. Behaviour can remain sensitive to which outcome previously followed which action, showing that the acquired relation can contain more structure than a generic habit of responding.

None of this requires one universal theory of representation.

The narrower conclusion is enough:

Conditioned behaviour can provide evidence of learning without uniquely revealing what was learned.

A behavioural response therefore answers one question better than another.

It can help establish that prior experience changed later behaviour.

It may require additional evidence to establish exactly which relation is controlling that behaviour.

Learned associations are expressed under conditions

An association is not a command that produces the same response whenever a cue appears.

A learned relation can influence behaviour differently depending on current circumstances.

Available actions matter. Motivation and bodily state can matter. Other learned relations can compete or combine. Current goals can alter how an outcome matters. Context can change which acquired relation is retrieved or expressed.

This continues the distinction established in D5.1: learning contributes to behaviour without uniquely determining behaviour.

The same acquired relation therefore need not be expressed identically across contexts.

How learned relations extend across similar cues and situations is developed more fully in Generalisation and Discrimination.

For D5.2, the foundational point is simply that associative expression is conditional.

Explicit report is another measure, not the definition

In humans, researchers can sometimes ask participants directly what relationship they noticed.

Those reports are valuable, but they are not identical to the associative effects measured through behaviour or physiology.

Some associative effects are closely related to explicit contingency awareness. Other findings show more complicated relationships between what people can report and how they respond.

The evidence does not support a simple universal rule that associative learning is either always conscious or always independent of awareness.

The safer distinction is:

Explicit knowledge about a contingency and behavioural expression of associative learning are related but not identical measures.

An arranged contingency is one thing. Evidence that a relation was learned is another. A person's explicit description of that relation provides an additional source of evidence.

D5.2 therefore does not define associative learning through conscious verbal knowledge, but neither does it treat association as a synonym for unconscious learning.

Association does not name one mechanism

The term associative learning can make the mechanism sound more settled than it is.

Researchers have explained associative phenomena using several kinds of models, including associative links; outcome expectancies; propositional representations; attentional processes; prediction-error updating; and hybrid accounts.

These theories often address the same behavioural phenomena from different explanatory perspectives.

No one of them should simply be substituted for the general concept.

Associative learning is a phenomenon to be explained, not the name of one universally accepted model of how learning occurs.

This distinction also changes how classic findings should be interpreted.

Cue competition shows why simple pairing accounts are insufficient, but it does not prove that one particular prediction-error algorithm explains all associative learning.

Evidence for outcome-specific learning challenges an exclusively generic stimulus–response account, but it does not prove that all associations must take the form of conscious propositions or explicit expectations.

Associative-learning experiments constrain theories.

They do not automatically settle the ontology of association itself.

Associative learning is not all learning

Associative learning occupies a major part of learning science, but it is not a synonym for learning.

D5.1 established a broader category of experience-dependent change. Habituation and sensitization, for example, are conventionally classified as forms of non-associative learning. Other forms of learning likewise need not fit neatly within the cue–outcome or action–outcome framework used here.

Associative learning is therefore a major class of learning, not an exhaustive definition of the phenomenon.

It is also not synonymous with habit formation. Associative processes can contribute to later habitual behaviour, but the relationship between learned actions, outcomes and habitual control requires further distinctions developed in What Makes Behaviour Habitual? and Goal-Directed and Habitual Control.

Later experience can also change associative responding in ways that require their own explanation. Extinction Is New Learning addresses one important case.

Two branches from here

Associative learning is often introduced as learning that two things "go together."

That description is too weak.

Experience can change sensitivity to relations among events, cues, actions and outcomes. Temporal proximity can matter, but contingency matters too. Prior learning can alter what is acquired from later experience. A relation present in the environment is not identical to the relation learned, and the relation learned is not identical to the behaviour through which researchers measure it.

This is the central architecture:

RELATION PRESENT IN EXPERIENCE → LEARNED RELATION → interacting with context, current conditions and other learning → OBSERVED EXPECTATION OR BEHAVIOUR

This figure is an analytic distinction, not a three-stage mechanism.

From here, associative learning separates into two major branches.

One asks how organisms learn relationships among cues and events.

That is the territory of Classical Conditioning.

The other asks how organisms learn relationships between their actions and subsequent outcomes.

That is the territory of Operant Learning.