Uninorm-Based Logic Neurons as Adaptive and Interpretable Processing Constructs

Abstract

Fuzzy logic neurons and networks as introduced more than a decade ago, seamlessly combine the transparency of logic and architectural underpinnings of neural networks. They dwell on logic connectives (operators) implemented in terms of t-norms and s-conorms along with some logic predicates of inclusion, dominance, similarity and difference. In this study, while adhering to the principle of fuzzy neural networks, we venture into the use of uninorms in place of triangular norms and co-norms. The paper offers new models of generic processing units, called AND and OR unineurons, discusses the underlying functionality of such neurons along with their numeric characteristics and develops comprehensive architectures of logic networks. Discussed is the use of Particle Swarm Optimization as a vehicle of the learning of the networks. Several illustrative numeric examples are included.