The neural mechanisms of worse than expected prediction errors

Learning about conditioned inhibitors, which predict omission of outcome delivery, has been relatively understudied compared to learning about reward predictors. Reward omissions lead to inhibition of dopamine neurons, driven by the lateral habenula, an important region that is also involved in learning about predictors of punishment. How could a conditioned inhibitor, which has no primary punishment association, learn to drive this dopamine dip signal? We show that the temporal-differences algorithm can account for learning negative associations for a conditioned inhibitor, and used this model to construct regressors for an fMRI conditioned inhibition experiment ran with 19 subjects. We found neural correlates of a prediction error for a CS in the ventral tegmental area, as well as value signals in the substantia nigra/VTA, along with the amygdala and caudate and pallidum regions of the basal ganglia. We also discuss a biologically based artificial neural network model called the PVLV (Primary Value, Learned Value) model, that can learn conditioned inhibition, and proposes that amygdala circuitry is involved in controlling dopamine firing for conditioned stimuli, while ventral striatal circuitry communicates with the lateral habenula to control dopamine firing for unconditioned stimuli. This model’s specification of the excitatory and inhibitory inputs to dopamine neurons in the different experimental conditions can be used to interpret fMRI signals in dopamine regions. This focus on worse than expected prediction errors can also be applied to understanding the learning about painful stimuli, and how disorders like depression and chronic pain involve distortions in these learning mechanisms leading to persistent negative predictions.