Distributed Estimation and Learning Under Communication and Privacy Constraints

Speaker Bios and Abstracts

• Bio

  Ayfer Ozgur received her B.Sc. degrees in electrical engineering and physics from Middle East Technical University, Turkey, in 2001 and the M.Sc. degree in communications from the same university in 2004. From 2001 to 2004, she worked as hardware engineer for the Defense Industries Development Institute in Turkey. She received her Ph.D. degree in 2009 from the Information Processing Group at EPFL, Switzerland. In 2010 and 2011, she was a post-doctoral scholar at the same institution. She is currently an Associate Professor in the Electrical Engineering Department at Stanford University where she is the Chambers Faculty Scholar in the School of Engineering. Dr. Ozgur received the EPFL Best Ph.D. Thesis Award in 2010, the NSF CAREER award in 2013, the Okawa Foundation Research Grant, Faculty Research Awards from Google and Facebook, the IEEE Communication Theory Technical Committee (CTTC) Early Achievement Award in 2018 and was selected as the inaugural Goldsmith Lecturer of the IEEE ITSoc in 2020.

• Abstract

  Motivated by the increasingly many applications of distributed and federated learning, in this talk we will consider the problem of learning high-dimensional distributions and their parameters from distributed samples under communication and local differential privacy constraints. We will develop novel encoding and estimation mechanisms that simultaneously achieve optimal privacy and communication efficiency in various canonical settings, including both probabilistic and so called “distribution-free” models. We will complement our achievability results with information-theoretic lower bounds based on data processing inequalities that describe how Fisher information from statistical samples scales with local differential privacy and communication constraints. We will compare the performance of digital and analog communication schemes in this context. We will conclude by discussing how our theoretical results apply to the distributed and federated training of deep neural network models and provide an experimental evaluation of our algorithms.