Item

Abstract

Decisions to avoid or escape predators occur at different spatiotemporal scales, resulting in different computations and neural circuits.
At their extremes, surprising or proximal threats will reduce decision and state space and utilize model-free architectures, while distant threats allow increased information processing supported by model-based operations.
Model-free and model-based computations, however, are often intertwined. Furthermore, under conditions of safety the foundations for effective reactive execution in the future can be laid through model-based instruction of model-free control. Prospective planning can also be enabled.
Together, these computations reflect distinct population codes embedded within a distributed defensive circuitry whose goal is to determine and realize the best policy.
Naturalistic observations show that decisions to avoid or escape predators occur at different spatiotemporal scales and that they are supported by different computations and neural circuits. At their extremes, proximal threats are addressed by a limited repertoire of reflexive and myopic actions, reflecting reduced decision and state spaces and model-free (MF) architectures. Conversely, distal threats allow increased information processing supported by model-based (MB) operations, including affective prospection, replay, and planning. However, MF and MB computations are often intertwined, and under conditions of safety the foundations for future effective reactive execution can be laid through MB instruction of MF control. Together, these computations are associated with distinct population codes embedded within a distributed defensive circuitry whose goal is to determine and realize the best policy.