Τίτλος: “Machine Learning and Machine Unlearning for Data Systems”.
Ομιλητής: Peter Triantafillou, Professor of Data Systems at the Department of Computer Science at the University of Warwick.
Ημερομηνία-χώρος: Παρασκευή 02 Δεκεμβρίου, 15:00, Αμφιθέατρο Γ, Τμήμα Μηχανικών Η/Υ και Πληροφορικής.
Περίληψη: Learned Data Systems (that is, data systems with machine learning components) bear the promise of increased performance, especially for resource-hungry analytics tasks over large datasets. As such, they are enjoying large attention by researchers. However, DB systems differ from other domains where machine learning (ML) plays a key role in that DBs are continuously updated. How can we ensure then that previously trained neural-network ML models continue to be accurate in the face of DB updates, such as data insert and/or delete operations? New data insertions may carry out-of-distribution (OOD) data for which models may be highly inaccurate. Likewise, for data deletions, which additionally introduce an additional challenge, namely that of unlearning. How can we then surgically unlearn what was previously learned and now deleted without erasing knowledge about relevant retained data? And how can we ensure the above efficiently, i.e., without retraining the models from scratch (which is a time-consuming operation)? In this talk, I will highlight our research results for the above problems. To our knowledge this is the first research results achieving the above goals.
Σχετικά με τον ομιλητή: Peter Triantafillou is Professor of Data Systems at the Department of Computer Science at the University of Warwick where he established and is currently leading the Data Sciences Theme. Peter is currently a Fellow of the Alan Turing Institute, a member of the Advisory Board of PVLDB, PC co-Chair of PVLDB Reproducibility, and Associate Editor for ACM SIGMOD 2022. Peter has served as a member of the Advisory Board of the Urban Big Data Research Centre (a UK national infrastructure for urban data services and analytics). Peter has previously held professorial positions at the University of Glasgow (UK), Simon Fraser University (Canada), Technical University of Crete and University of Patras (Greece) and visiting professorships at the Max-Planck Institute for Informatics (Germany).
Peter received his PhD in computer science from the University of Waterloo and was the Department of Computer Science and the Faculty of Mathematics nominee for the Gold Medal for outstanding achievements at the Doctoral level. Peter has published extensively in top journals and conferences in his areas, including 50+ papers in journal and 100+ papers in conferences and workshops. His papers have won numerous awards, including the most influential paper award in ACM DEBS 2019, the best paper award at the ACM SIGIR 2016 Conference, the best paper award at the ACM CIKM Conference 2006, and the best student paper award at IEEE Big Data 2018 Conference. Peter has served in the Technical Program Committees of more than 140 international conferences and has been the PC Chair or Vice-chair/Associate Editor in several prestigious conferences (including ACM SIGMOD, IEEE ICDE, IEEE DSAA, ACM Middleware, WISE, etc.). Finally, Peter is a co-designer of several innovative systems, such as the DBEst/DBEst++ engines for large-scale analytics, the MINERVA decentralized search engine, and the eXO decentralized social networking system.