Unified Neuro Symbolic Reasoning Algorithms Integrating First Order Logic With Transformers

Abstract

The integration of symbolic reasoning and deep learning is considered a revolutionary step in the field of artificial intelligence. This work proposes Unified Neuro-Symbolic Reasoning Algorithms (UNSRA), an innovative system that seamlessly combines first-order logic (FOL) and deep neural networks using transformer architectures. Conventional deep learning algorithms have great abilities to identify patterns but lack interpretability and logic-based reasoning, whereas classical symbolic AI approaches are capable of performing complex logic operations but cannot learn from raw data. This work takes advantage of both worlds by designing a two-path architecture in which neural networks deal with semantic representations and FOL components address logic-related tasks such as inference and rule implementation. The experimental results obtained using different datasets, such as Visual Question Answering, Knowledge Graph Completion, and Reasoning problems, show that the model outperforms existing neuro-symbolic models by 18% to 24% in logical consistency measures, while still competing effectively in benchmark tests. The findings of the interpretability study demonstrate that research have developed an approach that yields explainable decision-making processes that can be tracked from the neural layer to the symbolic layer, thus filling vital gaps in explainability in AI systems. Research contribution includes a flexible and modular system design that can be customized for different reasoning applications.