Neural-symbolic AI

Neural-symbolic AI integrates the learning power of deep neural networks with the interpretability and principled reasoning of symbolic AI systems. The goal is to overcome limitations of each approach: pure neural systems are powerful but opaque; pure symbolic systems are transparent but require hand-coded knowledge and struggle with complex, noisy real-world data.

Approaches to neural-symbolic integration include:

• Symbolic-guided neural learning: neural models trained with symbolic constraints or curriculum (e.g., logic rules regularizing loss)
• Neural reasoning over symbols: neural networks processing symbolic representations (e.g., knowledge graphs, logic predicates)
• Dual-system architectures: separate neural and symbolic subsystems that coordinate (e.g., neural encoding + symbolic reasoning)
• End-to-end neural learning of symbolic rules: neural systems that output interpretable symbolic representations (e.g., decision trees, logic rules) as part of their inference

Key papers

• TELLER: A Trustworthy Framework For Explainable, Generalizable and Controllable Fake News Detection — dual-system architecture where cognition system (neural, LLM-based) generates logic atoms from news text, and decision system (symbolic, disjunctive normal form layer) learns interpretable rules; enables both accuracy and human-understandable explanations
• [[2021-cincilioglu-neural-symbolic-rule-learning|Cincilioglu & Russo (2021) — ps2rule: End-to-end Neural-Symbolic Rule Learning]] — neural model that learns disjunctive normal form rules from data in an end-to-end fashion; applicable to classification tasks where interpretability is required

Connections

• Explainable AI — neural-symbolic systems are more interpretable by design than pure neural approaches
• Trustworthy AI — neural-symbolic methods support human oversight and intervention better than black-box neural models
• Fake news detection methods — emerging approach to balancing accuracy with explainability in fake news detection