Conference on Cognitive Computational Neuroscience

The ability to guide decisions based on context is a key building block of intelligent behavior. Here, we study the neural mechanisms underlying sequential decisions (plans) that are made in distinct task contexts. We asked participants in the fMRI scanner to navigate an avatar through a grid world to find rewards in two of four potential goal locations (rooms). Participants learned to perform the task in two distinct contexts, each of which defined the location of the second reward contingent on the first. Using multivariate analyses of BOLD signals, we asked how the neural geometry of spatial coding varied with goal. In the hippocampus, we found that the representation of space depended on context, with space being “compressed” so that context-specific goals were embedded together in neural space, and this compression predicted performance. A model in which place fields represent both current and prospective locations can account for these results.