CV

Education

• Ph.D in Explainable AI for Healthcare, Università della Svizzera Italiana, 2023 - Ongoing
  • Designing multimodal explainable Deep Learning models exploring causal models and counterfactuals.
• MPhil in Advanced Computer Science, University of Cambridge, 2022-2023
  • Grade: Distinction (highest grade possible)
• B.S. in IoT, Big Data & Web, Univestià degli studi di Udine, 2018-2021
  • Grade: 110/110 Cum Laude
  • The only student to complete it in two academic years rather than three

Work experience

• Research Intern @ MIT - Jul. 2023 - Aug. 2023
  • Selected among the 30 students to be part of SGI (Summer Geometric Initiative) organised by MIT.
  • Worked on 4 different research projects at the intersection between Geometric Processing and Deep Learning with mentors from MIT, Stanford and University of Toronto.
  • Supervisors: Dr Alberto Tono (Stanford University), Dr Peter Chen (MIT), Tongzhou Wang (MIT), Silvia Séllan (University of Toronto),
• Machine Learning Engineer @ Reply - Sep. 2021 - Sep. 2022
  • Spark and GCP were the main tools I used to build ML models when I worked as a consultant for one of the biggest telecom companies.
  • Developed ML models to maximize revenue and subscription, reducing customer churn. We improved revenues by $\sim$20%.
  • Completed the Databricks Certified Associate Developer for Apache Spark 3.0 course
• Research Assistant @ AILAB - Jul. 2020 - Jun. 2021
  • Conducted a research project about the usage of Graph Neural Network for Fake News Classification.
    I implemented a model similar to the SOTA on FEVER dataset and improved the previous results by $\sim$2% the Accuracy on HOVER dataset.
  • Analyzed data for a project about the retrieval of Adverse Drug Events related to Covid vaccines on Twitter. We also create a website.
  • Supervisors: Professor Giuseppe Serra (Università degli Studi di Udine), Dr. Enrico Santus (Bloomberg)

Projects

• Climbing the Ladder of Interpretability with Counterfactual Concept Bottleneck Models
  • First solution to integrate counterfactual in Concept-BasedModels.It does not need post-hoc methods, as it is trained end-to-end to predict, explain and generate counterfactuals.
  • It outperforms all the baselines in generating counterfactuals that are more accurate and feasible.
• SHARCS: SHARed Concept Space for Explainable Multimodal Learning - MPhil Project
  • First solution to use concepts to explain multimodal models. It is able to explain one modality using the other and retrieve a missing modality.
  • It performs as well as non-interpretable baselines, providing explanations and substantially better performance on missing modality situations (an increase of up to $\sim$84% accuracy)
• Effect of Aggregation and Layers in GNNs on Concept Finding - GNNs Project
  • Examined the effect of using different aggregators and GNNs on concept finding.
  • Proposed a new GNNs layer that used the similarity of the neighbourhood. It is theoretically more powerful than GCN and GAT and discovers better concepts.
• ProtoWNet - Weight prediction with prototype similarity - XAI project
  • Proposed a new version of ProtoPNet, that can discover a variable number of prototypes (taken from the train set) to solve the task.
  • It achieves comparable accuracy to the base model and better explanations (up to 5% improvement) than other methods on Node Classification tasks.
• Different ways to approach NER with BERT - NLP Project
  • Developed and compared different approaches to use BERT in WNUT-17 shared task. Restructured the input of BERT to approach the problem as Text Classification rather than Token Classification.
  • Achieved an improvement of 2.5% in the Macro F1 Score at word level.

Publications

SHARCS - Shared Concept Space for Explainable Multimodal Learning

Gabriele Dominici, Pietro Barbiero, Lucie Charlotte Magister, Pietro Liò, & Nikola Simidjievski. (2023). SHARCS: Shared Concept Space for Explainable Multimodal Learning.

Climbing the Ladder of Interpretability with Counterfactual Concept Bottleneck Models

Gabriele Dominici, Pietro Barbiero, Francesco Giannini, Martin Gjoreski, Marc Langheinrich & Giuseppe Marra. (2024). Climbing the Ladder of Interpretability with Counterfactual Concept Bottleneck Models

Teaching

Computer Architecture - Teaching Assistant

Graph Neural Network - Teaching Assistant

Awards

• Best graduate student of the year of my degree course - Università degli Studi di Udine
• Scholarship given to top-3 student of the degree course - Università degli Studi di Udine
• Best Storytelling Award and Best Engagement Award @ MeshUp Reloaded - Friuli Innovazione

Summer Schools

• DeepMind - Mediterranean Machine Learning Summer School 2022
• DeepMind - Eastern European Machine Learning Summer School 2021

Extracurricular Activity

• Mentee and Ambassador @ LeadTheFuture
  • Selected as one of the ambassadors of LeadTheFuture to share its mission with Italian students.
  • Among the few Italian students selected to be mentees for LeadTheFuture, a leading mentorship non-profit organization for students in STEM, with an acceptance rate below 20%.