Exploring the Brain Mechanisms Involved in Reward Prediction Errors Using Conditioned Inhibition

Many previous studies of the brain areas involved in reward prediction errors have not accounted for the downstream projections of dopamine areas when interpreting these results. We propose that paradigms like conditioned inhibition, which involves pairing a rewarded CS with an inhibitor that always cancels the reward, can reduce this confound and allow for further specification of the computational role different brain regions play into the RPE signal. Further predictions of the role of different brain areas in reward learning and how positive and negative valence learning interact in the brain are inspired by the PVLV model, a more biologically plausible alternative to TD learning that uses two parameters, learned value and primary value, compared to the single RPE parameter used by TD. To test the predictions of the PVLV model and compare activity across different conditions that allowed us to examine the roles of different regions in RPE computation, we ran a conditioned inhibition fMRI study using juice rewards, with a particular focus on examining the brain activity in several regions of interest in the PVLV model, including the ventral striatum, central nucleus of the amygdala and lateral habenula. We found that the rewarded CS activated the VTA/SN, consistent with many other studies of reward learning, as well as several regions in the basal ganglia. There was also overlap between activations for the CS, CS+Inhibitor and Inhibitor in the prefrontal cortex, insula and basal ganglia. Also, better than expected rewards activated the medial OFC; while worse than expected rewards activated the lateral OFC. In the amygdala, we found increased activity for the juice reward compared to the neutral solution, but we did not find increased activity for the rewarded CS as predicted by the PVLV model