2026-09-19T13:33:27Z https://www.ideals.illinois.edu/oai-pmh

oai:www.ideals.illinois.edu:2142/115797 2023-07-11 col_2142_5131 col_2142_10761 com_2142_5130 com_2142_10755 com_2142_234

Koyejo, Oluwasanmi 2022-05 application/pdf en text Submission original under an indefinite embargo labeled 'Open Access'. The submission was exported from vireo on 2022-11-11 without embargo terms The student, Shovik Guha, accepted the attached license on 2022-04-27 at 22:32. The student, Shovik Guha, submitted this Thesis for approval on 2022-04-27 at 22:37. This Thesis was approved for publication on 2022-04-29 at 08:38. DSpace SAF Submission Ingestion Package generated from Vireo submission #17990 on 2022-11-11 at 17:54:07 On sparse mirror descent Guha, Shovik 2022-04-29 Machine Learning Optimization Algorithms Mirror Descent Sparse Optimization Parsimony is a general guiding principle in science and philosophy which suggests that if one has multiple theories fitting the data equally well, one should choose the ``simplest" theory. In the field of machine learning and artificial intelligence, the sparsity of a model is used as a measure of parsimony. Algorithms which produce an optimal set of sparse parameters for a given model have been notoriously difficult to construct due to the non-convex and combinatorial nature of sparsity constraints. In this thesis we begin by giving an overview of popular algorithms for sparse and convex optimization. We then show how they can be combined with classical tools from the theory of approximation algorithms to compute approximate projections onto the sparsity constraints, which ultimately leads to a novel algorithm for sparse optimization. Thesis eng https://hdl.handle.net/2142/115797 Copyright 2022 Shovik Guha M.S. Thesis Computer Science University of Illinois at Urbana-Champaign Computer Science