Learning: How Predictions and Consequences Change Behaviour

Your phone sounds, and you anticipate a message before you look. You check it, find something interesting, and become more likely to check again. These are related but different learning problems: a signal predicts an event, while an action produces a consequence.

Learning concerns relatively lasting changes in knowledge or behavioural potential through experience. A temporary performance change caused only by fatigue or intoxication is not the same as learning. Nor does an absence of immediate performance prove that nothing was learned.

Learning goals and Feynman practice

Distinguish classical from operant conditioning, correctly label reinforcement and punishment, predict effects of reinforcement schedules, and explain why biology, information, and controllability constrain learning.

Read a section, close it, and teach the mechanism without relying on its name. Identify what predicts what, give a fresh example, and state one limitation. Repair the step you cannot explain, then attempt the hidden self-checks.

1. Classical conditioning: learning what a signal predicts

Pavlov investigated digestive physiology and studied how cues preceding food could come to elicit salivation. The important relationship is between events: a formerly neutral cue becomes informative about another stimulus.

Term	Meaning in this example	Example
Unconditioned stimulus, US	Elicits the relevant response without the experimental association.	Food.
Unconditioned response, UR	Response elicited by the US.	Salivation to food.
Conditioned stimulus, CS	Cue that acquires predictive significance through learning.	A tone preceding food.
Conditioned response, CR	Learned response elicited by the CS.	Salivation to the tone.

“Unconditioned” refers to the response relationship under study; it does not mean the organism has no developmental history. A conditioned response also need not be an exact copy of the unconditioned response. It can prepare the organism for what is expected.

A weather forecast is a useful analogy: its value lies in predicting what follows. The organism need not consciously describe the prediction, and physiological responses are not always chosen deliberately.

Timing, contingency, and surprise

Forward conditioning, with the CS preceding the US at a suitable interval, often supports learning. The appropriate interval depends on the response system. Taste-aversion learning, for example, can span long delays. Backward conditioning often produces different or weaker effects, but does not invariably produce no learning; a cue may signal the end of an event or a period without it.

Pairing alone is insufficient. A cue is less informative when the outcome occurs just as often without it. Rescorla’s experiments manipulated shock probability during and outside a cue, showing why the relationship between these probabilities matters. [1]

Consider a fictional cafeteria: if lunch is available only after a chime, the chime is useful. If lunch appears equally often with or without the chime, it adds little predictive information. These are teaching scenarios, not a reproduced dataset.

Prediction error is the difference between an expected and an obtained outcome in a learning model. Unexpected outcomes can update expectations more strongly than outcomes already fully predicted. This is more precise than saying that every repeated pairing stamps in an association equally.

Teach it back: How could two groups receive the same number of tone–food pairings but learn differently if one also receives food between tones?

2. Extinction, recovery, generalization, and discrimination

Extinction occurs when the CS repeatedly appears without the expected US and the conditioned response declines. It often involves new, context-sensitive learning rather than erasure of the original association.

Spontaneous recovery is a return of some responding after time has passed following extinction. Its occurrence supports distinguishing reduced performance from complete deletion. Recovery is not guaranteed on every test, and other forms of return can follow changes in context or outcome exposure.

Generalization is responding to cues similar to the trained CS. A dog trained with one tone may respond to nearby frequencies. The shape of the generalization gradient depends on learning and perceptual distinctions.

Discrimination training contrasts a CS+, associated with the outcome, with a CS−, not associated with it under the procedure. The organism can learn which cue predicts the event. A CS− is not automatically a conditioned inhibitor; additional evidence is needed for that stronger claim.

Higher-order conditioning occurs when a new cue acquires control over responding through pairing with an established CS, without direct pairing with the original US. Its strength and persistence depend on the procedure. It does not imply indefinitely strong chains of associations without direct outcomes.

Teach-back check: why does recovery after extinction matter?

It shows that loss of an observable response need not mean loss of every earlier association. The current context and competing learning can affect which response is expressed.

3. Conditioned emotion and physiological responses

Watson and Rayner’s Little Albert report is historically associated with attempts to condition fear. Its small, poorly controlled design, uncertain interpretation, and serious ethical problems mean it should not be treated as a clean demonstration that all phobias arise by the same mechanism.

Conditioning can influence physiological processes beyond obvious movement. Ader and Cohen paired a distinctive taste with an immunosuppressive drug and investigated later conditioned immune effects. [2] Appropriate unpaired controls and physiological measurements matter: avoiding a taste alone would not establish immune suppression. This research does not imply that thoughts can voluntarily replace medical treatment.

Teach it back: What would you need to measure to distinguish a learned dislike of a drink from a conditioned immune response?

4. Operant conditioning: learning what an action changes

In operant, or instrumental, conditioning, behaviour changes as a function of its consequences. The action need not involve a conscious theory of cause and effect. The central question is whether a consequence depends on the response and changes its future occurrence.

Thorndike studied cats escaping puzzle boxes and formulated the law of effect: consequences influence whether responses recur. Skinner developed controlled chambers and cumulative recording to study response patterns and their consequences systematically.

A cumulative record adds each response to a running total. A steep slope means a high response rate; a flat segment means a pause. The total does not fall when responding slows, so read its slope rather than treating it as a moment-by-moment rate graph.

Shaping versus chaining

Shaping reinforces successive approximations: first approaching a target, then touching it, then pressing it. As the criterion changes, earlier approximations alone no longer earn reinforcement.

Chaining links distinguishable actions into a sequence, such as collecting materials, setting up equipment, and completing a task. Shaping a response and chaining several responses are related techniques, not the same definition.

5. Four contingencies: two questions prevent most mistakes

Ask first: was something added or removed? Then ask: did the target behaviour increase or decrease? Positive means addition; negative means removal. Reinforcement increases behaviour; punishment decreases it.

Contingency	What changes?	Behavioural effect	Teaching example, if the effect occurs
Positive reinforcement	Something is added.	Behaviour increases.	Specific praise follows submitting work, and timely submission increases.
Negative reinforcement	Something is removed or avoided.	Behaviour increases.	Fastening a seat belt stops a buzzer, increasing prompt fastening.
Positive punishment	Something is added.	Behaviour decreases.	An extra correction task follows an error, and that error becomes less frequent.
Negative punishment	Something is removed.	Behaviour decreases.	Losing game access after a rule violation reduces the violation.

Classify by observed effects, not intention. A supposedly unpleasant reprimand might reinforce attention-seeking if it increases the behaviour. A reward that fails to increase the target response has not demonstrated reinforcement in that situation.

Negative reinforcement is not punishment. Relief can reinforce both useful and unhelpful behaviours: avoiding a difficult task may reduce anxiety immediately and make avoidance more likely later.

Teach-back check: a reminder stops when you submit a form, and submission becomes faster. What is the contingency?

Negative reinforcement of submission: an aversive reminder is removed contingent on the response, and the response increases or becomes more prompt. “Negative” describes removal, not whether the action is morally bad.

6. Reinforcement schedules: count responses or watch the clock?

Continuous reinforcement delivers reinforcement after each qualifying response. It is often useful for initial acquisition. Intermittent reinforcement reinforces some qualifying responses and can produce greater persistence when reinforcement later stops. This partial-reinforcement extinction effect depends on training and testing conditions; it is not a universal guarantee of slower acquisition or endless persistence.

Schedule	Rule	Typical pattern under suitable laboratory conditions
Fixed ratio, FR	Reinforce after a fixed number of responses.	A post-reinforcement pause followed by relatively rapid responding.
Variable ratio, VR	Reinforce after a varying response count around an average.	Relatively high, steady responding.
Fixed interval, FI	Reinforce the first qualifying response after a fixed interval.	Often a pause followed by accelerating responding.
Variable interval, VI	Reinforce the first qualifying response after varying intervals.	Relatively steady responding.

An interval schedule is not simply delivering a reward after time passes regardless of behaviour. That would be a time-based, noncontingent schedule. The response requirement is essential.

Variable-ratio examples such as some gambling arrangements illustrate uncertain response-based outcomes, but real machines can involve probability-based schedules and additional design features. The example is explanatory, not advice to use gambling-like incentives.

Teach it back: Why would pressing faster increase rewards under some ratio schedules but not produce unlimited rewards under an interval schedule?

Self-check: specify the contingency

A ratio schedule ties reinforcement to response count. An interval schedule requires time to elapse before a response can produce reinforcement, so extra responses before availability do not each earn a reward.

7. Dopamine: learning signals are not simply pleasure

Midbrain dopamine systems contribute to reinforcement learning, motivation, and action selection. In influential models, some dopamine responses reflect a reward prediction error: an unexpectedly good outcome produces a different signal from a fully predicted outcome, and omission can also matter.

Different dopamine neurons and circuits have different functions. Dopamine is not the only learning mechanism, nor does every dopamine signal measure pleasure. Operant learning also involves representations of actions, outcomes, context, and bodily state across broader networks.

Teach it back: Why should a fully expected reward teach less about a prediction than a surprising one, even if both are enjoyable?

8. Biological preparedness: associations are not equally easy

Garcia and Koelling’s experiments showed selective associations between cues and consequences: taste was especially effective in learning about illness, while audiovisual cues were more effective with external shock under their procedures. [3]

This challenges the idea that any arbitrary cue associates equally readily with any outcome. Taste aversion can also arise after very few experiences and long cue–illness delays.

The conclusion is relative preparedness, not a permanent inability to learn every other association. Species, exposure, timing, and procedures matter. Likewise, evolved threat sensitivity may contribute to some fears, but does not show that phobias inevitably target ancient dangers or cannot involve modern objects.

Teach-back check: why does selective learning challenge a simple contiguity-only theory?

If temporal pairing alone determined learning, matched pairings should have similar effects regardless of cue–outcome combination. Selectivity shows that the organism’s sensory and biological organization also constrains what is learned.

9. Learned helplessness: controllability changes later behaviour

Classic experiments associated prior uncontrollable aversive exposure with reduced later escape behaviour. The traditional account proposed learning that actions do not affect outcomes. An important comparison used groups receiving similar aversive exposure but differing in whether their responses controlled it, helping separate controllability from exposure alone.

Later neuroscience refined that interpretation. Maier and Seligman emphasized that prolonged aversive stimulation can produce passivity through default defensive circuitry, while learning control engages systems that can counteract those effects. [4] Present the original explanation as historically important rather than the final settled mechanism.

This animal research is not a diagnosis of an individual who feels discouraged, nor does it establish that human depression has one cause. The general research question is how experience of control influences future action and stress responses.

Teach it back: Why compare controllable and uncontrollable conditions with similar exposure instead of comparing only exposure with no exposure?

10. Tolman: learning can exist before its expression

In latent-learning studies, rats explored mazes without an explicit food reward. Introducing a reward later could produce a rapid improvement, consistent with useful learning during earlier exploration.

The distinction is between acquisition, what was learned, and performance, what the animal expresses under the current motivation and task. A poor initial score need not mean no knowledge was acquired.

Cognitive maps describe representations of spatial relations that can support flexible navigation. The results challenge accounts requiring an explicit food reinforcer for every learning event; they do not prove that every rat acquired a complete map or that all improvement occurred instantly under every procedure. Exploration also contains sensory and motivational consequences, so “no explicit food reward” is more precise than “no consequences of any kind.”

An everyday example is learning a building’s layout while wandering, then using that knowledge when you later need an exit. This is an analogy, not an equivalent experimental test.

11. Worked example: why do you keep checking your phone?

A notification sound can become a conditioned cue predicting a message. Anticipatory arousal is different from the checking action. If checking sometimes produces valuable information and becomes more frequent, a consequence may be reinforcing that response. If checking removes uncertainty or anxiety, relief could also reinforce it.

Classify the mechanism by what depends on what. A notification that arrives independently of your behaviour is not automatically an operant reward for checking, and irregular messages do not automatically establish a variable-ratio schedule. Define the actual response–outcome rule before naming a schedule.

Teach it back: Identify a cue–outcome association and an action–consequence relationship in a different everyday routine.

Retrieval workshop and spaced review

From memory, label US, UR, CS, and CR in a fresh example. Reconstruct the four-contingency table and four schedules. Explain why extinction differs from erasure, why preparedness differs from impossibility, and why performance differs from learning.

Repair: “A consequence is reinforcement because it feels pleasant”

It is reinforcement when its contingent delivery increases the relevant behaviour. Pleasantness may help explain effectiveness but does not substitute for observing the behavioural change.

Apply: design a test of predictive value rather than pairings alone

Give two groups comparable cue–outcome pairings, but add outcomes outside the cue in one group. Compare later cue-elicited responses while controlling exposure and context. This tests whether the cue adds information beyond the background outcome rate.

Review tomorrow and in a week. Score each explanation: 0 for a label, 1 for a definition, 2 for a mechanism plus example and limitation. Repair your weakest explanation before adding more terms.

One-minute explanation: Classical conditioning links signals with events; operant conditioning links actions with consequences. Effective learning depends on prediction, contingency, timing, and biological organization. Reinforcement and punishment are defined by changes in behaviour, while schedules determine when consequences follow responses. Learning may remain latent until conditions support its expression, and control over outcomes can alter future behaviour.

Sources and further reading

The supplied numbered citations did not include a bibliography. These references support the prediction, immune-conditioning, preparedness, and controllability examples. Everyday scenarios and schedule illustrations are teaching aids.

1. Rescorla, R. A. (1968). Probability of shock in the presence and absence of CS in fear conditioning.
2. Ader, R., and Cohen, N. (1975). Behaviorally conditioned immunosuppression.
3. Garcia, J., and Koelling, R. A. (1966). Relation of cue to consequence in avoidance learning.
4. Maier, S. F., and Seligman, M. E. P. (2016). Learned helplessness at fifty: insights from neuroscience.