A0529
Title: Investigating the RGB approach for safe AI
Authors: Emanuela Raffinetti - University of Pavia (Italy) [presenting]
Paolo Giudici - University of Pavia (Italy)
Golnoosh Babaei - University of Pavia (Italy)
Abstract: Artificial intelligence applications require the development of practical tools that can mitigate risks arising from their use. Specifically, in order to ensure the trustworthiness of artificial intelligence systems, the four main key principles of sustainability (robustness), accuracy, fairness, and explainability have to be achieved. Recently, a new methodology named ''Rank graduation box'' (RGB) was introduced as a unified approach which shares the same theoretical root and allows to overcome one of the main drawbacks of the existing methods, i.e. the high computational effort. The behavior of the RGB metrics is further explored by means of simulation experiments, which can be easily reproduced. The experimental results indicate that these metrics are easy to interpret and that they can be applied to any machine-learning model independently of the underlying data.
