Responsabile: Mario Pasquato
Descrizione: Variable X-ray sources observed by the XMM-Netwon telescope and curated by the Extras project contain a wealth of astronomical information which is accessible only through machine learning approaches due to the large amount of data, which cannot be inspected manually.
We are currently using supervised classification methods to identify certain sources of interest. Unfortunately, most machine learning methods (e.g. boosted trees, neural nets) are not interpretable: we do not know why a classifier casts its prediction for a given data point.
The candidate will explore two approaches to alleviate this problem: eXplainable AI (XAI) methods, which produce post-hoc explanations for a black box classifier; and inherently interpretable models. The student will become familiar with machine learning in Python (using scikit-learn and possibly neural networks in tensorflow or pytorch) and apply XAI tools and interpretable models via Python modules. They will assess which tool is most adequate for our purposes in terms of speed, reliability, and of providing insight into the classification process. Interpretability is an exploding field in machine learning, together with the neighboring fields of physics informed learning and algorithmic fairness, endowing the candidate with transferable skills in high demand in the job market.
Durata: 9-12 mesi
Laurea: Magistrale
Prerequisiti: Nessuno