BEGIN:VCALENDAR VERSION:2.0 PRODID:-//Memento EPFL// BEGIN:VEVENT SUMMARY:IEM Distinguished Lecturers Seminar: Kernel-driven and Learnable S elf-Supervision over Graphs DTSTART:20240920T131500 DTEND:20240920T140000 DTSTAMP:20260509T195820Z UID:8043ce08bef5eae1cb2ebff34884abac5792817a3b2a9cc8722c0048 CATEGORIES:Conferences - Seminars DESCRIPTION:Prof. Georgios B. Giannakis\, University of Minnesota\, USA\n* ** NEW PLACE: BM 5202 ***\n\nCoffee and cookies will be served at 13:00 be fore the seminar \n\nAbstract \nSelf-supervision (SeSu) has gained popular ity for “data-hungry” training of machine learning models\, especially those involving large-scale graphs\, where labeled samples are scarce or unavailable. Main learning tasks in such setups are ill-posed\, and SeSu r enders them well-posed by relying on abundant unlabeled data as input\, to yield low-dimensional embeddings of a reference (auxiliary) model output. In this talk\, we first outline SeSu approaches\, specialized reference m odels\, and their links with (variational) auto-encoders\, regularization\ , semi-supervised\, transfer\, meta\, and multi-view learning\; but also t heir challenges and opportunities when multi-layer graph topologies and mu lti-view data are present\, when nodal features are absent\, and when the ad hoc selection of a reference model yields embeddings not optimally desi gned for the downstream main learning task. Next\, we present our novel Se Su approach which selects the reference model to output either a prescribe d kernel or a learnable weighted superposition of kernels from a prescribe d dictionary. As a result\, the learned embeddings offer a novel\, reduced dimensionality estimate of the basis kernel\, and thus an efficient param etric estimate of the main learning function at hand that belongs to a rep roducing kernel Hilbert space. If time allows\, we will also cover online variants for dynamic settings\, and regret analysis founded on the so-term ed neural-tangent-kernel framework to assess how effectively the learned e mbeddings approximate the underlying optimal kernel(s). We will wrap up wi th numerical tests using synthetic and real datasets to showcase the merit s of kernel-driven and learnable (KeLe) SeSu relative to alternatives. The real data will also compare KeLe-SeSu with auto-encoders and graph neural networks (GNNs)\, and further test KeLe-SeSu on reference maps with maske d inputs and predicted-outputs that are popular in large language models ( LLMs).\n\nBio\nGeorgios B. GIANNAKIS is a Professor of Electrical and Comp uter Engineering at the University of Minnesota\, where he holds a Preside ntial Chair. His interests span the areas of statistical learning\, commun ications\, and networking - subjects on which he has published over 495 jo urnal papers\, 805 conference papers\, 26 book chapters\, two edited books \, and two research monographs. His current research focuses on Data Scien ce with applications to IoT\, and power networks with renewables. He is th e (co-) inventor of 36 issued patents\, and the (co-)recipient of 10 best journal paper awards from the IEEE Signal Processing (SP) and Communicatio ns Societies\, including the G. Marconi Prize. He received the IEEE-SPS No rbert Wiener Society Award (2019)\; EURASIP's A. Papoulis Society Award (2 020)\; Technical Achievement Awards from the IEEE-SPS (2000) and from EURA SIP (2005)\; the IEEE ComSoc Education Award (2019)\; and the IEEE Fourier Technical Field Award (2015). He is a member of the Academia Europaea\, G reece's Academy of Athens\, and Fellow of the National Academy of Inventor s\, the European Academy of Sciences\, UK's Royal Academy of Engineering\, Life Fellow of IEEE\, and EURASIP. He has served the IEEE in several post s\, including as a Distinguished Lecturer for the IEEE-SPS. LOCATION:BM 5202 https://plan.epfl.ch/?room==BM%205202 https://epfl.zoom.u s/j/67449285011 STATUS:CONFIRMED END:VEVENT END:VCALENDAR