[Theory] REMINDER: Talks at TTIC: 9/16 Davoud Ataee Tarzanagh, University of Michigan

Brandie Jones bjones at ttic.edu
Thu Sep 15 08:00:00 CDT 2022


*When:   *         Friday, September 16th at* 10:30am CT *

*Where:*         Talk will be given live, *in-person* at
                         TTIC, 6045 S. Kenwood Avenue
                         5th Floor, Room 530

*Virtually:  *    via Panopto (Join Here
<https://uchicago.hosted.panopto.com/Panopto/Pages/Viewer.aspx?id=10474b07-61b0-4b60-b14b-af0f00f7cd9d>
)

*Who:              *Davoud Ataee Tarzanagh, University of Michigan

*Title:              *Federated Bilevel Optimization and its Applications
in Machine Learning

*Abstract:     *Federated learning (FL) is a distributed learning approach
that enables model training on a large collection of decentralized data.
Existing FL methods are widely applied to stochastic problems with
single-level structure. However, many contemporary machine learning
problems - including adversarial robustness, hyperparameter tuning,
actor-critic, and neural architecture search - fall under nested bilevel
programming that subsumes minimax and compositional optimization. Towards
addressing such nested problems, bilevel optimization has received
significant attention in the recent literature, albeit in non-FL settings.
On the other hand, federated versions have been elusive perhaps due to the
additional challenges surrounding heterogeneity, communication, and inverse
Hessian approximation.


In this talk, we present a federated machinery for general bilevel problems
with lightweight communication. We provide convergence rates for the
proposed approach in the presence of heterogeneous data and introduce
variations for minimax and compositional optimization problems. Under
certain regularity conditions, applying our results to these problems
either improves or matches the best-known sample complexity in the non-FL
settings. We complement our theory with experiments on fair classification
and hyper-representation learning that demonstrate the benefits of our
method in practice.


-- 
*Brandie Jones *
*Administrative Assistant*
Toyota Technological Institute
6045 S. Kenwood Avenue
Chicago, IL  60637
www.ttic.edu

Working Remote on Tuesdays
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cs.uchicago.edu/pipermail/theory/attachments/20220915/e9b6bb8c/attachment.html>


More information about the Theory mailing list