Again, this comparison revealed no significant difference (p = 0.14, t = 1.55). Taken together, these analyses suggest that the representation of sensory evidence as well as unspecific BOLD responses in early sensory areas did not change significantly over the course of learning. So far we have shown that (1) the predictions IWR-1 cost of an adapted reinforcement learning model correlate with learning-related changes in orientation discrimination performance over time and (2) that the model-derived DV, which builds the basis for perceptual
decisions, is coded in the medial frontal cortex. However, because alternative learning models would also predict similar increases in DV over learning, in the following analyses we provide further evidence for the proposed reinforcement learning mechanism. Evidence for Rescorla-Wagner-like updating in the reward-learning literature originally came from Z-VAD-FMK in vitro the observation of signed reward prediction error signals in dopamine neurons ( Bayer and
Glimcher, 2005 and Schultz et al., 1997). In human fMRI studies, however, prediction error signals have been identified in the ventral striatum, a target area of dopaminergic midbrain neurons ( Kahnt et al., 2009, McClure et al., 2003, O’Doherty et al., 2003 and Pessiglione et al., 2006). Thus, to provide further evidence for a reinforcement learning process in the current perceptual learning task, we regressed the signed prediction errors from the model against the feedback-locked BOLD signal in each voxel (see Experimental Procedures). We identified significant (p < 0.0001, k = 5) correlations between model-derived prediction errors and activity in the left ventral striatum ([-9, 0, −3],
t = 4.77; ADP ribosylation factor Figure 7A), the bilateral anterior insular cortex extending into the lateral OFC (left BA 47 [-33, 21, −3], t = 5.56; right BA 47 [30, 21, −6], t = 6.49), the dorsolateral PFC (right BA 9 [54, 15, 36], t = 5.17), as well as the dorsomedial prefrontal cortex including the ACC (BA 32 [0, 27, 42], t = 5.81; Figure 7B; see Table S3 for complete results). This shows that the key learning variable of our computational model, namely the signed reward prediction error, is coded in the activity of reward-related regions such as the ventral striatum, providing further evidence for a reinforcement learning process in perceptual learning. In a second step, we aimed to confirm that the learning-related changes in DV are indeed related to an updating mechanism that is based on signed prediction errors as proposed by our model. Thus, the same region in the ACC where activity patterns track perceptual learning-related changes in DV should also process reward prediction error signals.