´Ü´Ç´Ç³¾ÌýÌý
Speaker
Andrea Bertozzi Ìý| University of California
Title
Geometric Graph-Based Methods for High Dimensional Data
Ìý´¡²ú²õ³Ù°ù²¹³¦³Ù
High dimensional data can be organized on a similarity graph - anÌýundirected graph with edge weights that measure the similarity between data assigned to nodes.ÌýWe consider problems in semi-supervised and unsupervised machine learningÌýthat are formulated as penalized graph cut problems. There are a wideÌýrange of problems including Cheeger cuts, modularity optimization onÌýnetworks, and semi-supervised learning.ÌýWe show a parallel between these modern problems and classical minimalÌýsurface problems in Euclidean space. Ìý
This analogy allows us to develop a suite of new algorithms for machine learning that are both very fast andÌýhighly accurate. ÌýThese are analogues of well-known pseudo-spectralÌýmethods for partial differential equations.