ETD

• explainability

The thesis tackles two problems in the recently-born field of Explainable AI (XAI),
and proposes some algorithms to solve them. XAI has the overarching goal of
providing human-understandable explanations of Machine Learning models, which,
nowadays, operate as highly complex black-box models whose decisions, especially in
high-stakes and critical settings, we are not able to understand.
The thesis tackles the novel problem of Local-to-Global (L2G) explainability, and
local explainability. In a L2G setting one wishes to infer an understanding of the
overall behavior of a model starting from explanations of its punctual decisions, that
is, to infer global explanations from local ones. We propose two Local-to-Global
algorithms to tackle this problem, Rule Relevance Score and GLocalX.
Then, we focus on local explainability, and provide an algorithm, TriplEx, to
explain Transformer-based models on a variety of tasks.