Research map / Large language models / AI and society
From Statistical Relational to Neuro-Symbolic Artificial Intelligence
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Survey that positions ~20+ neurosymbolic (NeSy) and statistical relational AI (StarAI) systems along seven shared dimensions (inference, syntax, semantics, learning scope, representation, paradigms, tasks) to foster cross-fertilization between the two fields.
Problem
The two communities integrating logic with learning (neurosymbolic AI and statistical relational AI) have developed largely separately, with little shared vocabulary or comparison, making systems hard to characterize and compare.
Approach
Identifies seven dimensions shared by NeSy and StarAI (proof vs model-based inference, logical syntax, semantics, parameter vs structure learning, symbolic vs subsymbolic representation, paradigm integration, and learning tasks) and uses them to categorize systems and draw parallels between them.
Key results
- Positions logic-based NeSy frameworks (e.g., DeepProbLog, LTN, SBR, Semantic Loss, DeepStochLog, Scallop) along six dimensions in a summary table.
- Argues that proof-based NeSy systems build neural architectures from proof trees while model-based systems encode logic as loss functions, with a control-versus-scalability tradeoff.
- Shows via examples that fuzzy relaxations with a 0.5 threshold can yield Boolean assignments violating the original formula, and that fuzzy and probabilistic semantics disagree on idempotent conjunctions.
Surprising findings
- Naive fuzzy relaxation of a disjunction can have a global fuzzy optimum (0.34, 0.34, 0.34) that discretizes to an assignment violating the original Boolean formula.
- Under the product t-norm, the fuzzy score of a2 <- b,b is 0.09 while a1 <- b is 0.3, whereas probabilistic semantics assigns both 0.3.
Limitations
- Focuses on a logical perspective and representative systems rather than complete coverage of the field.
- Survey does not run experiments; comparisons are conceptual and tabular.
Experimental setup
Builds on: ProbLog Markov Logic Networks DeepProbLog Semantic Loss Knowledge-Based Artificial Neural Networks
Closest papers by meaning
- LogiCity: Advancing Neuro-Symbolic AI with Abstract Urban Simulation (NeurIPS 2024 · similarity 0.86)
- Relational Neurosymbolic Markov Models (AAAI 2025 · similarity 0.85)
- EM-NeSy: Expectation Maximization for Neurosymbolic Learning (NeurIPS 2026 · similarity 0.85)
- NeSyCoCo: A Neuro-Symbolic Concept Composer for Compositional Generalization (AAAI 2025 · similarity 0.84)
- Efficient Rectification of Neuro-Symbolic Reasoning Inconsistencies by Abductive Reflection (AAAI 2025 · similarity 0.84)
- A learnability analysis on neuro-symbolic learning (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.