A model of schema learning based on biological dimensionality reduction during sleep

Reorganizing learned knowledge into generalized representations and transferring it to future learning are essential aspects of cognition, often described as schema formation and use. However, the computational and circuit mechanisms underlying these processes remain unclear. Here, we propose a theoretical model in which schemas emerge through the formation and alignment of low-dimensional neural representations. In this model, high-dimensional input patterns are reorganized into low-dimensional manifolds through replay-driven Hebbian nonlinear dimensionality reduction. Manifold alignment simultaneously maps representations with shared task structure onto a common format, enabling downstream readout circuits to be reused across tasks. The model captures three core features of schema learning: rapid learning of similar tasks by reusing low-dimensional representations from prior experience, sleep-dependent generalization to unobserved relationships in transitive inference, and compositional recombination of schemas to solve novel tasks. Together, these results suggest a potential neural mechanism for forming, aligning, and recombining low-dimensional schemas to support future learning.