Compositional Neuro-Symbolic Concepts in Neural Activities

ICLR 2026 Conference Submission14108 Authors

Keywords: concept grounding, neural decoding, neuro-symbolic systems

TL;DR: A neuro-symbolic framework for grounding compositional concepts in fMRI responses

Abstract: We explore whether human neural responses exhibit compositional structure via NEURONA, a modular neuro-symbolic framework for grounding compositional concepts in neural activity. Leveraging image- and video-based fMRI question-answering datasets, NEURONA learns to map interacting concepts from visual input to patterns of fMRI signals, explicitly modeling their relational structure through hierarchical predicate-argument dependencies. We demonstrate that incorporating these structural priors improves both decoding accuracy and generalization to unseen visual stimuli. Our findings provide support that relational meaning is better explained by guided co-activation across multiple regions, and highlight neuro-symbolic frameworks as promising tools for decoding compositional concepts from neural activity.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 14108