Schema-based active inference supports rapid generalization of experience and frontal cortical coding of abstract structure

A team led by Dileep George from Vicarious AI, along with researchers from Ghent University and Italy's National Research Council, has developed a groundbreaking computational framework called Schema-based Hierarchical Active Inference (S-HAI). This model combines principles from predictive processing and active inference with schema-based mechanisms to explain how biological brains rapidly generalize knowledge from limited experiences. The framework features a two-level architecture where a higher-level generative model encodes abstract task structure while a lower-level model handles spatial navigation, connected through a grounding likelihood function that maps abstract goals to physical locations.

Through detailed simulations, S-HAI demonstrated remarkable capabilities in reproducing human-like rapid generalization behaviors. The model showed the ability to flexibly remap abstract schemas to novel contexts, resolve goal ambiguity, and balance between reusing existing knowledge versus accommodating new information. Crucially, the framework's neural activity patterns matched experimental observations from rodent medial prefrontal cortex during schema-dependent navigation tasks, including task-invariant goal-progress cells and goal-spatially conjunctive cells.

This research provides the first comprehensive mechanistic account that bridges behavioral patterns, neural coding data, and computational theory of schema-based learning. The findings suggest that the brain's remarkable generalization abilities may emerge from predictive processing principles implemented hierarchically across cortical and hippocampal circuits. This work represents a significant step toward understanding the fundamental computational principles underlying intelligent behavior in both biological and artificial systems.