Semantic Artificial Neural Networks

Abstract

Neural networks have achieved in recent years human level performance in various application domains, including critical applications where accountability is a very important issue, closely related to the interpretability of neural networks and artificial intelligence in general. In this work, an approach for defining the structure of neural networks based on the conceptualisation and semantics of the application domain is proposed. The proposed approach, called Semantic Artificial Neural Networks, allows dealing with the problem of interpretability and also the definition of the structure of neural networks. In addition, the resulting neural networks are sparser and have fewer parameters than typical neural networks, while achieving high performance.