Aplicando XAI na comparação de redes neurais e árvores de decisão

Abstract

The new focus on artificial intelligence confronts us with a long-standing concern data on such algorithms, the distinction between white-box and black box. This study seeks to explore analysis techniques, known as Explainable AI (XAI) for such models, especially in so-called black-box algorithms, those in which the details of his decision-making are not completely known. Is important the explainability of the AI model, as opaque models can covertly inflict ethical and trustworthiness issues, including the possibility of bias, discrimination, privacy and rights violations. The chosen approach aims to study and apply some of these XAI techniques to unravel the logic of an algorithm considered black-box, using a progressive approach, which explores the fundamentals of a neural network through to application of explainable techniques, presenting a comparison of behavior between neural networks and decision tree. Finally, comparisons are made between metrics of the models and the results found are discussed. ​