Glossary
Prediction error
What is Prediction error?
A prediction error is a difference between an expected and an observed outcome. It is a learning signal in many computational models, including models of reward learning.
How it works
Models use prediction errors to revise estimates. In reward models, an outcome better than expected gives a positive error and a worse outcome a negative error. Primate dopamine findings helped connect these computational ideas to neural responses; they do not establish one universal signal for all learning.
What the primate recordings showed
Schultz, Dayan and Montague describe tasks in which monkeys received rewards such as fruit juice and learned that a preceding cue predicted them. Recordings from midbrain dopamine neurons showed a characteristic sequence:
- Unexpected reward: a brief increase in firing followed the reward.
- Learned prediction: the brief increase shifted to the predictive cue, while the fully expected reward no longer produced that extra burst.
- Expected reward absent: firing fell below its baseline near the time the reward should have arrived.
The same desirable outcome could therefore produce different activity depending on its predictability. These recordings support a reward-prediction-error interpretation for the neurons and tasks studied. Absence of a reward-related burst does not mean all neural activity or learning has stopped.
Prediction error example
You expect an ordinary dinner and get an excellent one. That illustrates a positive reward prediction error: the outcome exceeded the expectation. It does not, by itself, show what your dopamine neurons did.
Why it matters
Prediction errors make expectations and updating explicit and testable. Their form and biological implementation depend on the model and task.
Source: Schultz, Dayan and Montague (1997) connect reward prediction with primate dopamine findings.