Neural and Computational Mechanisms of Reward and Aversion

Mollick, Jessica

Graduate Thesis Or Dissertation

Neural and Computational Mechanisms of Reward and Aversion Public Deposited

Analytics

Citations

Citeable URL: https://scholar.colorado.edu/concern/graduate_thesis_or_dissertations/x633f1154

Abstract

While many fMRI studies have focused on understanding which brain areas encode reward prediction errors, interpretation of these signals is difficult given that dopamine neurons also encode reward prediction errors, and send widespread modulatory signals throughout the brain. Using conditioned inhibition, where participants learn about a conditioned inhibitor that always cancels the expected reward, reduces these effects of dopamine, as dopamine neurons respond with an inhibition, or dopamine dip, to a conditioned inhibitor. These dip signals are driven by a brain region called the lateral habenula. In the conditioned inhibition fMRI study, we found evidence that the habenula responds to a conditioned inhibitor, and replicated prior studies showing activity for reward predictors in midbrain dopamine regions (SN/VTA).Next, using a biological model of the dopamine system, the PVLV model, with different pathways for dopamine for a conditioned and unconditioned stimulus, and positive and negative valence learning, we test hypotheses about the sources of the BOLD signal. An analysis on the fMRI task found that the predicted absolute value of the dopamine from the model, meaning that an expected inhibition of dopamine neurons was converted to an increase in BOLD, was a good fit for the BOLD signals in midbrain dopamine regions, as well as downstream regions like the nucleus accumbens that receive dopamine. The absolute value of predicted dopamine in a pain learning task was also correlated with BOLD signals in the substantia nigra (SN), amygdala and nucleus accumbens.Another region involved in aversive learning is the PAG (periaqueductal gray), and we suggest that it reduces responses to an aversive US based on the level of expectation, rather than encoding a full prediction error signal. In the pain task, we found evidence for the PAG encoding a modulatory signal, and an expectation signal in the amygdala. This approach can help clarify some of the mixed findings on the reward prediction error encoding properties of the dopamine system and its role in reward and punishment learning. It can be applied to cases where risky decisions to pursue positive outcomes continue to be made in the face of potential aversive consequences.

Creator