Evidence for Model-based Computations in the Human Amygdala during Pavlovian Conditioning

Contemporary computational accounts of instrumental conditioning have emphasized a role for a model-based system in which values are computed with reference to a rich model of the structure of the world, and a model-free system in which values are updated without encoding such structure. Much less studied is the possibility of a similar distinction operating at the level of Pavlovian conditioning. In the present study, we scanned human participants while they participated in a Pavlovian conditioning task with a simple structure while measuring activity in the human amygdala using a high-resolution fMRI protocol. After fitting a model-based algorithm and a variety of model-free algorithms to the fMRI data, we found evidence for the superiority of a model-based algorithm in accounting for activity in the amygdala compared to the model-free counterparts. These findings support an important role for model-based algorithms in describing the processes underpinning Pavlovian conditioning, as well as providing evidence of a role for the human amygdala in model-based inference. A hot topic in the neurobiology of learning is the idea that there may be two distinct mechanisms for learning in the brain: a model-based learning system in which predictions are made with respect to a rich internal model of the learning environment, versus a “model-free” mechanism in which trial-and-error learning occurs without any rich internal representation of the world. While the focus in the literature to date has been on the role of these mechanisms in instrumental conditioning, almost nothing is known about whether more fundamental kinds of learning such as Pavlovian conditioning also involve model-based processes. Furthermore, nothing is known about the extent to which the amygdala, which is known to be a core structure for Pavlovian learning, contains neural signals consistent with a model-based mechanism. To address this question, we used a novel Pavlovian conditioning task and scanned human volunteers with a special high-resolution fMRI sequence that enabled us to obtain signals within the amygdala with over four times the resolution of conventional imaging protocols. Using this approach in combination with sophisticated computational analyses, we find evidence to suggest that the human amygdala is involved in model-based computations during Pavlovian conditioning.