For cognitive control of behavioral adjustments in feedback learning, various processings are required, including evaluating the saliency (i.e., relevance to the task) and hedonic value of feedback information for future response selection. In this study, brain regions involved in processing feedback saliency were investigated by comparing activations for positive feedback (following correct responses) and negative feedback (following errors) for early and late phases of learning. A conditional associative learning task was used in which stimulus-response association rules were learned by trial and error, based on the feedback. Since there were four available responses to choose among for each stimulus, only positive feedback (i.e., reward) was relevant to behavioral adjustment during the early learning phase of learning, but negative feedback (e.g., penalty) became more relavant as learning progressed. fMRI data obtained from normal adults (n = 29) were analyzed to identify brain regions where responses to each feedback varied across the four consecutive runs. Activation for reward decreased as learning progressed, whereas activation for penalty increased in the following areas: anterior insula, dmPFC and anterior cingulate region, inferior PFC, inferior parietal cortex, and cerebellum. We interpret these results as reflecting the decreased saliency of positive feedback and increased saliency of negative feedback, between early and late phases of the learning task. In contrast, for two areas associated with processing of hedonic value, the ventral striatum and vmPFC, activations (positive > negative feedback) did not vary across the four consecutive runs. These observations suggest that the saliency of feedback for learning is processed in a network separate from that for the hedonic value of feedback.