%0 PDF
%T Incremental dictionary learning via geometric sparse encoding
%A Hudes, Matthew I.
%8 2023-05-19
%E Tasissa, Abiy Fekadu
%I Tufts Archival Research Center
%R http://localhost/files/m900p8808
%X We consider the incremental learning of sparse representations of high dimensional data, whereby learning occurs continuously from a stream of data. This is in contrast to learning methods that assume all data can be accessed at once. We propose a framework based on a geometric regularizer that encourages sparsity by representing data points via local landmark points (atoms). The sparse representations and the atoms can be learned from a set of points using alternating minimization. To test the effectiveness of the sparse regularizer, we use a simple prototype whereby atoms are vertices of a Delaunay triangulation and data points are sampled from each simplex. Using the proposed framework, we design multiple incremental learning algorithms which we test on a subset of the MNIST database of handwritten digits.; Thesis (B.S.)--Tufts University, 2023.; Submitted to the Dept. of Mathematics.; Advisor: Abiy Tasissa.
%[ 2023-06-29
%9 http://purl.org/dc/dcmitype/Text
%~ Tufts Digital Library
%W Institution