Influence of Inherent Values on Associative Learning

In an initial study we used high and low facial attractiveness as a proxy for inherent value and associated each face either with a high or a low probability of winning a small amount of money. When we analyzed the data we found an clear effect of value congruence: very attractive faces that predicted a high probability of winning were selected most frequently, whereas unattractive faces that predicted a high probability of losing were selected least frequently. Furthermore, reaction times to value congruent choices (either high/high or low/low) steadily decreased over time indicating that these cues processed with decreasing effort. For value incongruent trials, reaction times remained the same. Computational modeling revealed that value congruence affected the speed of learning (learning rate) with which the new expected values were learned: learning rates for congruent trials were initially higher (indicating faster initial learning), whereas learning rates for incongruent trials were very small indicating that the new expected values were learned as a slower speed. This suggests that value incongruence lead to more effortful and slow learning of new values that contradict the inherent values of the cue. The fMRI data revealed a hitherto novel neural correlated for a dynamically changing learning rate in the dorsal striatum in close proximity of other striatal regions that encoded important decision-related computations (e.g. prediction errors, reward and punishments). Furthermore, a connectivity analysis demonstrated that these different regions interact in much the same way that out reinforcement learning model updated the expected values.

Value Congruence and Contingency Reversals

In two follow-up studies we are investigating if and how value congruence effects change in face of a reversal con reward contingencies. After new expected values have been learned through cue-outcome associations, we changed the reward probabilities, such that a cue that was previously predicting a high probability of winning is now predicting a high probability of losing. We are interested in whether the value congruence effects on learning rates are reset after such a reversal, whether they are reinstantiated, but at a lower level, or whether learning keep constantly declining. Preliminary findings indicate that learning rates are reset after a reversal, but the magnitude of the initial learning rates after reversal remain complex. Furthermore, in these studies we also switched to cues of emotionally positive and negative valence and to cues of congruent and incongruent political statements. We are therefore also testing the generality of the previously observed value congruence effect with different inherent values.