# A Mathematical Framework, a Taxonomy of Modeling Paradigms, and a Suite of Learning Techniques for Neural-Symbolic Systems

**Authors:** Charles Dickens, Connor Pryor, Changyu Gao, Alon Albalak, Eriq Augustine, William Wang, Stephen Wright, Lise Getoor  
**arXiv:** [2407.09693](https://arxiv.org/abs/2407.09693)  
**Date:** 2024-07-12  
**Summary page:** https://www.summarizepaper.com/en/paper/2407.09693v1/

## Key points

- Rapidly evolving field of Neural-Symbolic (NeSy) systems
- Need for a unifying theory to elucidate commonalities and differences among NeSy systems
- Introduction of NeSy-EBMs as a mathematical framework for discriminative and generative modeling
- Development of taxonomy of modeling paradigms focusing on neural-symbolic interface and reasoning capabilities
- Suite of learning techniques tailored for NeSy-EBMs, including four distinct approaches leveraging methodologies from various domains
- Introduction of an open-source NeSy-EBM library designed for scalability and expressivity
- Extensive empirical analysis showcasing practical advantages of NeSy-EBMs in tasks such as image classification, graph node labeling, autonomous vehicle situation awareness, and question answering
- Establishment of a unifying framework connecting Neural-Symbolic AI with energy-based modeling through bilevel optimization techniques

## Summary

, , , , 

In the rapidly evolving field of Neural-Symbolic (NeSy) systems, there is a growing interest in achieving symbiotic unions of neural and symbolic methods. However, each NeSy system varies significantly in its fundamental approach, highlighting the need for a unifying theory to elucidate commonalities and differences among these systems. This paper introduces , which serve as a comprehensive mathematical framework for discriminative and generative modeling within both probabilistic and non-probabilistic NeSy approaches. Through the utilization of NeSy-EBMs, the authors develop a taxonomy of modeling paradigms that specifically focuses on a system's neural-symbolic interface and reasoning capabilities. Additionally, this paper presents a suite of learning techniques tailored for NeSy-EBMs, enabling the derivation of general expressions for gradients of prominent learning losses. This includes four distinct learning approaches that leverage methodologies from various domains such as bilevel and stochastic policy optimization. Furthermore, the paper introduces , an open-source NeSy-EBM library designed for scalability and expressivity to facilitate real-world applications of NeSy systems. Extensive empirical analysis conducted across multiple datasets showcases the practical advantages of NeSy-EBMs in diverse tasks including image classification, graph node labeling, autonomous vehicle situation awareness, and question answering. Building upon existing research on integrating symbolic knowledge with neural networks, this work establishes a unifying framework by connecting Neural-Symbolic AI with energy-based modeling. By employing bilevel optimization techniques, a new family of algorithms is proposed for end-to-end integration of symbolic reasoning with deep learning architectures. This comprehensive approach contributes significantly to bridging the gap between neural networks and symbolic reasoning systems within machine learning literature.