Hierarchy in neuronal representations of multiple tasks in prefrontal cortex

Abstract

The ability to perform multiple tasks is one of the fundamental hallmarks of general intelligence in the brain. However, the underlying neural mechanisms remain largely unknown. We trained macaque monkeys to perform four tasks with identical spatial layouts but different cognitive demands, and recorded activities of thousands of prefrontal neurons using two-photon calcium imaging. Within the same neuronal population, we identified a multitask neural geometry composed of multiple low-dimensional subtask spaces, each encoding spatial information within a specific subtask. These subtask spaces shared a ring-like representational structure, forming a generalized spatial code across different subtasks. Task separation arose from separable bases and offsets of subtask spaces, modulated primarily by meta-task features. Thus, the neural geometry is organized by a three-level hierarchical structure: location codes are nested within subtask spaces, which are grouped as different meta-tasks. This hierarchy supported generalization across both locations and tasks, and explained monkeys error behaviors. Together, the hierarchically organized neural geometry underlies flexible multitask behavior.