BEGIN:VCALENDAR VERSION:2.0 PRODID:researchseminars.org CALSCALE:GREGORIAN X-WR-CALNAME:researchseminars.org BEGIN:VEVENT SUMMARY:Gitta Kutyniok (Ludwig-Maximilians-Universität München) DTSTART:20210311T170000Z DTEND:20210311T180000Z DTSTAMP:20260423T035625Z UID:Inverse/37 DESCRIPTION:Title: Graph Convolutional Neural Networks: The Mystery of Generalization\nb y Gitta Kutyniok (Ludwig-Maximilians-Universität München) as part of Int ernational Zoom Inverse Problems Seminar\, UC Irvine\n\n\nAbstract\nThe tr emendous importance of graph structured data due to\nrecommender systems o r social networks led to the introduction of\ngraph convolutional neural n etworks (GCN). Those split into spatial\nand spectral GCNs\, where in the later case filters are defined as\nelementwise multiplication in the frequ ency domain of a graph.\nSince often the dataset consists of signals defin ed on many\ndifferent graphs\, the trained network should generalize to si gnals\non graphs unseen in the training set. One instance of this problem\ nis the transferability of a GCN\, which refers to the condition that\na s ingle filter or the entire network have similar repercussions on\nboth gra phs\, if two graphs describe the same phenomenon. However\,\nfor a long ti me it was believed that spectral filters are not\ntransferable.\n\nIn this talk we aim at debunking this common misconception by\nshowing that if tw o graphs discretize the same continuous metric\nspace\, then a spectral fi lter or GCN has approximately the same\nrepercussion on both graphs. Our a nalysis also accounts for large\ngraph perturbations as well as allows gra phs to have completely\ndifferent dimensions and topologies\, only requiri ng that both\ngraphs discretize the same underlying continuous space. Nume rical\nresults then even imply that spectral GCNs are superior to spatial\ nGCNs if the dataset consists of signals defined on many different\ngraphs .\n LOCATION:https://researchseminars.org/talk/Inverse/37/ END:VEVENT END:VCALENDAR