Logic-oriented fuzzy neural networks: A survey

• The work presents a comprehensive review and discussion of logic-oriented fuzzy AND\OR network. • The architecture exhibits outstanding learning and interpretation capabilities. • Logic-oriented neurons are endowed with clearly delineated logic structure. • Interpretability is the main advantage of the neurons that make the model linguistically readable. • Knowledge is extracted in the form that is easily comprehended by humans. Data analysis and their thorough interpretation have posed a substantial challenge in the era of big data due to increasingly complex data structures and their sheer volumes. The black-box nature of neural networks may omit important information about why certain predictions have been made which makes it difficult to ground the reliability of a prediction despite tremendous successes of machine learning models. Therefore, the need for reliable decision-making processes stresses the significance of interpretable models that eliminate uncertainty, supporting explainability while maintaining high generalization capabilities. Logic-oriented fuzzy neural networks are capable to cope with a fundamental challenge of fuzzy system modeling. They strike a sound balance between accuracy and interpretability because of the underlying features of the network components and their logic-oriented characteristics. In this survey, we conduct a comprehensive review of logic-oriented fuzzy neural networks with a special attention being directed to AND\OR architecture. The architectures under review have shown promising results, as reported in the literature, especially when extracting useful knowledge through building experimentally justifiable models. Those models show balance between accuracy and interpretability because of the prefect integration between the merits of neural networks and fuzzy logic which has led to reliable decision-making processes. The survey discusses logic-oriented networks from different perspectives and mainly focuses on the augmentation of interpretation through vast array of learning abilities. This work is significantly important due to the lack to similar survey in the literature that discusses this particular architecture in depth. Finally, we stress that the architecture could offer a novel promising processing environment if they are integrated with other fuzzy tools which we have discussed thoroughly in this paper.