Rescorla-Wagner Algorithm

• Rescorla & Wagner (1972): animals and humans also learn associations by paying Attention to what is not associated.
• $\Delta V=\alpha {\beta }_1{\beta }_2(\lambda =\Sigma V)$
• ▶ V = association strength ▶ ∆ V : Change in association strength ▶ λ = maximum values of the unconditional stimulus ▶ Set to 1: when US is present (food) ▶ Set to 0: when not present ▶ α = learning rate ▶ β = varies the effects of negative or positive evidence ▶ ΣV = sum of associated strengths for all cues/Features/conditions stimuli
• negative instances are also useful to learning
• Logical Problem of Lang Acquisition
• Children don’t get negative evidence = must be innate
• Cross-situational learning
• Propose-but-verify
• Rescorla-Wagner Blocking
• Rescorla-Wagner = error-driven
• After a strong association is made, as long as it is confirmed by data, no new learning will occur
• The model only learns when the predicted outcome differs from actual outcome