Rescorla–Wagner: learning = prediction error (Rescorla & Wagner)

ΔV = α · β · (λ − ΣV)

• ΔV = change in associative strength this trial · α, β = salience/learning-rate of the cue and the outcome · λ = the maximum association the outcome can support (what actually happened) · ΣV = the total already predicted by all present cues. The term (λ − ΣV) is the prediction error — the surprise.

Difference form = you only learn from what you didn't expect. When the outcome is fully predicted (ΣV = λ), the error is zero and nothing is learned — the famous consequence: a fully-expected event, however important, updates nothing (blocking). Surprise is the entire currency of learning.

Transfer. The academic root of "刺激 = 偏离基线 = prediction error" (two-pleasures-arousal-and-closure) and of von-restorff-effect / novelty. A message that only confirms what the audience already predicts (the median take) produces ΔV ≈ 0 — literally forgettable. This is anti-mid stated as a learning rule: to be encoded at all, an artifact must carry a term the audience's model did not already predict.

Kin: illustrative-formulas · anti-mid · von-restorff-effect · berlyne-arousal-curve · two-pleasures-arousal-and-closure.

Source: Robert Rescorla & Allan Wagner, 'A Theory of Pavlovian Conditioning: Variations in the Effectiveness of Reinforcement and Nonreinforcement,' in Classical Conditioning II (1972) — the prediction-error learning rule