Neural Networks in View of Explainable Artificial Intelligence

Abstract

Artificial neural networks are just mathematical functions used for approximation purposes, and learning is a euphemism for determining the parameters of these functions. Experience from other areas of approximation theory shows that closed-form approximation functions determined globally such as neural networks tend to be quite rippled and are, thus, outperformed by approximation functions defined locally. With respect to artificial intelligence, neural networks lack explainability. Therefore, research in novel approaches for local approximators apt to replace neural networks and to outperform them with respect to accuracy, computational expense and explainability is suggested.