Neurosymbolic Approaches in AI Design – An overview

Artificial Intelligence (AI) has seen exponential growth over the last few decades, due to significant advancements in neural networks and other intricate machine learning models. One significant challenge ahead is reconciling the interpretability of symbolic AI with neural networks. The key goal is to integrate neural networks with symbolic frameworks to enable reasoning, explainability, and logical pathways within innovative, complementary architectures that mitigate the shortcomings of both neural networks and symbolic AI. This research investigates NS-AI approaches in the past decade through a systematic review, adhering to a set of criteria created to answer the most fundamental of questions on designing an NS-AI approach. The objective of this study is to unveil a generic standardized design, documenting the integration of NS-AI approaches within a model based on neural networks and symbolic AI. The output of this research shows three clear differentiations in the structure of NS-AI approaches, which are named as follows: Sequential, Multi-Integration and Hybrid. The differentiation on the design phase of the approach brings a more coherent perspective overall and promotes a better understanding of an approach, ascertaining to which of the NS-AI structures it falls into based on its fundamental design.