Outcome uncertainty increases generalization in human predictive learning via prediction error

Uncertainty refers to the state in which the probability of an outcome following an event is unknown. Substantial progress has been made in terms of understanding how uncertainty modulates learning and decision making, however the effect of outcome uncertainty on the generalization of prior knowledge to novel situations has not been clearly documented. In two predictive learning experiments, participants experienced cue-outcome trials in which the probability that an outcome would follow a Gabor grating (vertical or horizontal) was deterministic (100%) or probabilistic (50% or 25%). Participants learned that one Gabor orientation predicted the outcome whereas the other did not. Following training, participants provided expectancy ratings of the outcome to the trained and to novel gratings at different orientations (generalization test). Experiment 1 documented that participants displayed more generalization to novel gratings when training was probabilistic. Experiment 2 replicated these results and further showed broader generalization when participants were trained with deterministic outcomes and then switched to probabilistic in the last training block, a manipulation that increased experienced prediction error and hence surprise. These results suggest that outcome uncertainty broadens generalization gradients, and this is mediated by unsigned prediction errors.