Answers are grounded in this paper and cite their sources.
Neuro-symbolic systems aim to combine neural learning with symbolic reasoning for building interpretable and generalizable intelligent systems. While recent studies have explored different strategies for neuro-symbolic integration, existing approaches remain limited by (i) the absence of unified learning mechanisms between continuous neural representations and discrete symbolic concepts, (ii) weak interaction between neural prediction and symbolic reasoning, and (iii) restricted reuse of learned concepts for unseen reasoning scenarios. To address these issues, we propose NSF-SRL, a general neuro-symbolic framework based on statistical relational learning. NSF-SRL jointly models neural predictions, latent predicates and first-order logic rules within a unified probabilistic framework. For concept learning, NSF-SRL introduces a neuralized variational inference mechanism that learns predicate-level representations from neural representations and enables efficient neuro-symbolic interaction through variational expectation-maximization optimization. For concept manipulation, NSF-SRL allows learned concepts to be reused and composed with logic rules for interpretable explanation and compositional reasoning. Extensive experiments on digit image addition, zero-shot image classification and visual relationship detection, together with comprehensive analyses, demonstrate the effectiveness, generalization and scalability of NSF-SRL.
No takes yet. Share an insight, caveat, or question.
Yu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: