Variable Influence aNd Contribution: a methodology for Explainable AI in Neural Networks

Explainability in Artificial Intelligence is a recent and important field of research and application. The powerfulness of AI models needs to meet the human necessity to understand how a decision was made. For this reason, many explainability methodologies are being developed. In this dissertation, a new methodology over neural networks is developed and compared to the already existing ones. It is called VINC, which stands for Variable Influence aNd Contribution, and it is able to separate in a precise way each feature contribution throughout the neural network until the final result. In this dissertation are discussed the theoretical and practical aspects of VINC in its research context and with some application examples. After evaluating the results, the impact, the best use cases and the future development of VINC are taken into consideration. ​