<div class="csl-bib-body">
<div class="csl-entry">Leger, F., & Aubin, P.-C. (2024). Extending convexity and gradient descent: a framework for general costs based on alternating minimization. In <i>Journees SMAI MODE</i> (pp. 15–15). http://hdl.handle.net/20.500.12708/198947</div>
</div>
-
dc.identifier.uri
http://hdl.handle.net/20.500.12708/198947
-
dc.description
Références
[1] F. Léger and P.-C. Aubin-Frankowski. Gradient descent with a general cost.
arxiv.org/abs/2305.04917, 2023.
[2] I. Csiszár and G. Tusnády. Information Geometry and Alternating Minimization Procedures. Statistics and Decisions, 205–237, 1984.
[3] H. H. Bauschke, J. Bolte and M. Teboulle. A descent lemma beyond Lipschitz gradient
continuity: first-order methods revisited and applications. Mathematics of Operations Research, 42(2):330–348, 2017
-
dc.description.abstract
In this talk we go beyond the quadratic cost in optimization, by replacing it with a
general cost function and using a majorize-minimize framework. We unveiled in [1] a new class of gradient-type optimization methods that extends vanilla gradient descent, mirror descent, Riemannian gradient descent, and natural gradient descent, while keeping the same proof ideas and (sub)linear rates of convergence. Our approach involves constructing a surrogate for the
objective function f in a systematic manner, based on a chosen cost function c. This surrogate ϕ(x, y) = c(x, y)+fc
(y) is then minimized using an alternating minimization scheme. We obtain
novel rates of convergence for this general algorithm based on Csiszar and Tusnady’s “five-points
inequality” [2].
Specifying the results to cost functions c, using optimal transport theory we establish convergence rates based on generalized notions of L-smoothness and convexity, namely c-concavity and the c-cross-convexity that we introduce. We provide local versions of these two notions when the cost satisfies a condition known as nonnegative cross-curvature. In particular our framework
encompasses the concept of relative smoothness introduced in 2017 in [3] for Bregman divergences, and we provide the first global rates for natural gradient descent and Newton’s method under a new assumption distinct from self-concordance.
en
dc.description.sponsorship
FWF - Österr. Wissenschaftsfonds
-
dc.language.iso
en
-
dc.subject
continuous optimization
en
dc.subject
Gradient Descent
en
dc.subject
General Cost
en
dc.title
Extending convexity and gradient descent: a framework for general costs based on alternating minimization
en
dc.type
Inproceedings
en
dc.type
Konferenzbeitrag
de
dc.contributor.affiliation
Institut national de recherche en informatique et en automatique, France
-
dc.description.startpage
15
-
dc.description.endpage
15
-
dc.relation.grantno
P 36344N
-
dc.type.category
Abstract Book Contribution
-
tuw.booktitle
Journees SMAI MODE
-
tuw.relation.publisherplace
Lyon
-
tuw.project.title
Unilateralität und Asymmetrie in der Variationsanalyse
-
tuw.researchTopic.id
C4
-
tuw.researchTopic.name
Mathematical and Algorithmic Foundations
-
tuw.researchTopic.value
100
-
tuw.publication.orgunit
E105-04 - Forschungsbereich Variationsrechnung, Dynamische Systeme und Operations Research
-
dc.description.numberOfPages
1
-
tuw.event.name
Journées SMAI MODE (2024)
fr
tuw.event.startdate
27-03-2024
-
tuw.event.enddate
29-03-2024
-
tuw.event.online
On Site
-
tuw.event.type
Event for scientific audience
-
tuw.event.place
Lyon
-
tuw.event.country
FR
-
tuw.event.institution
SMAI
-
tuw.event.presenter
Aubin, Pierre-Cyril
-
tuw.event.track
Single Track
-
wb.sciencebranch
Mathematik
-
wb.sciencebranch.oefos
1010
-
wb.sciencebranch.value
100
-
item.openairecristype
http://purl.org/coar/resource_type/c_5794
-
item.grantfulltext
restricted
-
item.openairetype
conference paper
-
item.cerifentitytype
Publications
-
item.fulltext
no Fulltext
-
item.languageiso639-1
en
-
crisitem.author.dept
Institut national de recherche en informatique et en automatique
-
crisitem.author.dept
E105-04 - Forschungsbereich Variationsrechnung, Dynamische Systeme und Operations Research
-
crisitem.author.parentorg
E105 - Institut für Stochastik und Wirtschaftsmathematik