# Domain-specific schema reuse supports flexible learning to learn in the primate brain

## TL;DR

The primate brain solves the stability-plasticity dilemma by reusing shared knowledge while keeping it separate from task-specific changes.

## Contribution

The study shows that the brain represents shared and task-specific information in near-orthogonal subspaces to enable flexible learning.

## Key findings

- The decision subspace contains neural correlates of schema that are reused across tasks.
- The decision and stimulus subspaces are nearly orthogonal, reducing interference between domains.
- This separation allows the brain to maintain stability while adapting to new environments.

## Abstract

Prior knowledge accelerates subsequent learning of similarly structured problems – a phenomenon termed learning to learn – by forming generalizable neural representations called neural correlates of schema (NCS). However, how the brain exploits stable NCS while remaining flexible towards changes (the stability-plasticity dilemma) remains unclear. Here, we show that the primate brain addresses this dilemma by representing the stable NCS and task-unique changes in a near-orthogonal manner. We analyzed neural activities in the dorsolateral premotor cortex of three male macaques trained to perform a series of visuomotor mapping tasks. By delineating decision and stimulus-related subspaces, we identified NCS within the decision subspace, whose reuse facilitated subsequent learning. In addition, the decision subspace exhibited a near-orthogonal relationship with the stimulus-related subspace, minimizing cross-domain interference. Our results reveal that restricting NCS to specific functional domains can preserve useful knowledge while maintaining near-orthogonality with other subspaces, enabling flexible adaptation to new environments, thereby resolving the stability-plasticity dilemma.

This study reveals that during a series of learning tasks, the macaque brain represents the shared (e.g. the decision strategy) and task-unique information (e.g., sensory inputs) in a near-orthogonal manner, thereby supporting reuse of the shared component to facilitate learning, while remain flexible to changes in specifics.