Neural-Symbolic AI: A Comprehensive Survey on Integrating Neural Learning with Symbolic Reasoning for Explainable and Generalizable Intelligence

The limitations of standalone neural networks such as lack of interpretability and symbolic systems like poor scalability to noisy data have spurred the development of neural-symbolic AI (NeSy AI), a hybrid paradigm that merges data-driven learning with explicit logical reasoning to achieve more transparent, robust, and generalizable intelligence. This survey synthesizes advancements from ۲۰۲۰ to ۲۰۲۵, including logic-constrained transformers and knowledge-infused large language models, as documented in venues like NeurIPS and ACM Computing Surveys. We outline foundational concepts, from symbolic knowledge graphs to neural architectures like CNNs and Transformers, and trace NeSy's evolution from early neuro-logic hybrids to differentiable theorem provers. A structured taxonomy categorizes approaches: embedding-based for relational integration, sequential-integrated hybrids, neuro-symbolic Transformers for linguistic tasks, and logic-guided methods for constrained optimization. Applications in explainable diagnostics (e.g., medical imaging with ۱۵-۲۵% accuracy gains in low-data regimes), robotics planning, cybersecurity, and reasoning-enhanced NLP are analyzed, emphasizing metrics like fidelity (up to ۹۰% in benchmarks) and zero-shot generalization. Challenges, including gradient flow across discrete-continuous boundaries and dataset curation, are critiqued, with proposals for scalable benchmarks and automated rule extraction. By consolidating studies, this review advances NeSy toward ethical AI and AGI pathways, fostering human-centric systems.

کلیدواژه ها:

Neural-Symbolic AI ، Explainable AI ، Differentiable Reasoning ، Knowledge-Infused Neural Models ، Logic-Constrained Learning ، Generalization