UC Berkeley

Studies of learning in both biological and artificial agents have shown that trial-and-error learning is capable of much complex behavior, but is inflexible and does not cope well with novel situations. This requires planning and the capacity to simulate the consequences of actions, but how this is accomplished remains unclear. Here, we recorded from hundreds of neurons simultaneously from the orbitofrontal cortex (OFC) and the hippocampus (HPC) while monkeys performed a navigation task to uncover the processes underlying planning. When planning their route, animals would frequently look at future places along their planned route. The HPC represents a cognitive map during navigation that is consistent with the functional structure of the environment, and this representation is reactivated during planning. In contrast, neuronal dynamics in the OFC alternate between two orthogonal value codes, one for the value of the current location and one for the future location to which the animal was looking. The OFC value code would sometimes switch to the future location even though the animal was not looking at a future location consistent with them thinking about the future even though behavior was not changing. By encoding actions and plans in two orthogonal neural subspaces, the OFC can maintain information about the organism's current state while simulating potential future states.