Classical Conditioning: Learning Predictive Relationships
Error-correction learning updates associative strength according to prediction error.
What Is Left to Learn?
Suppose a cue already predicts an outcome. If the same outcome continues to occur, does the learner gain as much from each new pairing as it did from the first one? The Rescorla-Wagner model answers no. Learning depends on the difference between what the learner predicts and what actually happens. When the outcome is surprising, associative strength can change substantially. When the outcome is already predicted, little new learning is needed.
What do you think happens?
A cue already predicts the outcome accurately. What happens when that outcome occurs again?
Reveal answer
Answer: The prediction error is low, so little additional learning occurs.
The model treats learning as error correction. If the existing prediction already accounts for the outcome, there is little mismatch left to correct.
Error-Driven Updating
The Rescorla-Wagner model describes learning as an adjustment to associative strength based on prediction error. Associative strength represents how strongly a cue contributes to predicting an outcome. Prediction error is the mismatch between the actual outcome and the total outcome predicted by the cues present. A large mismatch provides a basis for a larger update. A small mismatch provides a basis for a smaller update. If the prediction already accounts for the outcome, the error can be zero or low, and little further learning occurs.
The important comparison is not simply whether an outcome occurs. The comparison is between the outcome that occurs and the total prediction generated by the cues present. Learning is therefore sensitive to surprise. A cue associated with an unexpected outcome has more error to correct than a cue present when the outcome was already fully predicted.
Prediction Error Across Trials
A Cue Becomes Predictive
Trace what happens when a cue initially provides little prediction and later provides a strong prediction of the outcome.
Early pairing: At the beginning, the cue does not account for much of the outcome. The actual outcome differs substantially from the prediction, so the prediction error is relatively large and the cue has a basis for a noticeable associative-strength update.
Developing prediction: As associative strength changes, the cue accounts for more of the outcome. The difference between the prediction and the actual outcome becomes smaller.
Established prediction: When the cue already accounts for the outcome, the prediction error can be low or zero. The model therefore predicts little additional learning from another outcome that is already expected.
Learning is strongest when the outcome is not yet well predicted and becomes smaller as the prediction increasingly matches the actual outcome.
This progression does not mean that the cue stops being useful. It means that the cue has already learned much of what the outcome can teach under the current prediction. The remaining error, rather than the mere repetition of the pairing, determines whether substantial additional updating is justified.
Why Blocking Occurs
Blocking occurs when prior conditioning makes one cue a strong predictor of the outcome, and a second cue is then added to form a compound conditioned stimulus. Because the original cue already predicts the outcome, the compound produces little or no new prediction error when the outcome occurs. The added cue is present, but the outcome is already accounted for by prior learning. The Rescorla-Wagner model therefore gives the new component little or no additional associative strength.
Original Cue and Added Cue
An original cue has already been conditioned to predict an outcome. A new cue is then presented together with the original cue, followed by the same outcome. Why does the new cue acquire little associative strength?
Before the compound: Prior conditioning has made the original cue predictive of the outcome.
Compound presentation: The original cue and the added cue appear together. The outcome still occurs, but the original cue already accounts for it.
Error assessment: Because the total prediction is already close to the actual outcome, the prediction error is zero or low.
Learning about the added cue: With little prediction error available, the rule has little reason to increase the added cue's associative strength.
The added cue is blocked because prior learning already predicts the outcome; the issue is low prediction error, not that the added cue must be ignored perceptually.
Dividing Predictive Strength
In a compound conditioned stimulus, more than one cue is present, but the relevant prediction is the total prediction produced by the cues together. If the original cue already supplies the needed prediction, the added cue receives little additional associative strength. This is why associative strengths should be understood in relation to the compound's total prediction rather than by asking only whether each cue was physically present.
Connection to Least Mean Square
The Rescorla-Wagner model is closely related to the Least Mean Square learning rule. Both treat learning as an adjustment of a cue's weight or associative strength in response to prediction error. The adjustment is directed toward reducing the mismatch between the predicted outcome and the actual outcome. In the Rescorla-Wagner vocabulary, associative strengths are updated. In the Least Mean Square vocabulary, weights are adjusted. The shared idea is error-driven correction.
| Learning framework | What is adjusted? | What drives adjustment? |
|---|---|---|
| Rescorla-Wagner model | Associative strength | Prediction error |
| Least Mean Square learning rule | Weight | Prediction error |
Common Reasoning Errors
Assuming that every cue paired with an outcome must learn strongly.
The model bases learning on prediction error, not simply on the presence of a cue during the outcome.
Fix:
Check whether the total prediction already accounts for the outcome. If it does, the new cue receives little or no additional associative strength.Explaining blocking by saying that the added cue was necessarily ignored.
The source explanation focuses on low prediction error, not on a required failure to perceive the added component.
Fix:
Say that prior conditioning already predicted the outcome, leaving little new surprise for the added cue to explain.Treating repetition alone as the cause of learning.
When the prediction already matches the outcome, repetition produces little additional error-correction learning.
Fix:
Ask how much mismatch remains between the prediction and the actual outcome.Treating Rescorla-Wagner and Least Mean Square as unrelated ideas.
Both frameworks describe adjustments driven by prediction error.
Fix:
Map associative strength to weight and compare their shared error-reduction logic.
Apply the Error Test
A cue has already become a reliable predictor of an outcome. A second cue is then added, and the same outcome follows. Explain whether the second cue should gain substantial associative strength. Your answer should identify the total prediction, the prediction error, and the consequence for learning about the second cue.
Hints
- Start with what the original cue predicts before the second cue is added.
- Compare the total prediction from the compound with the actual outcome.
- Use the size of the remaining prediction error to explain the amount of learning.
Key Takeaways
- The Rescorla-Wagner model treats learning as error correction based on prediction error.
- A larger mismatch between the actual outcome and the total prediction creates a stronger basis for updating associative strength.
- As a cue's prediction increasingly matches the outcome, prediction error becomes smaller and additional learning decreases.
- Blocking occurs when prior conditioning already predicts the outcome, leaving little error for a newly added cue to explain.
- Rescorla-Wagner associative-strength updates and Least Mean Square weight updates share an error-reduction principle.