Class Central is learner-supported. When you buy through links on our site, we may earn an affiliate commission.

YouTube

Interaction Is Necessary for Distributed Learning with Privacy or Communication Constraints

Association for Computing Machinery (ACM) via YouTube

Overview

Save Big on Coursera Plus. 7,000+ courses at $160 off. Limited Time Only!
Explore a conference talk that delves into the necessity of interaction for distributed learning under privacy or communication constraints. Examine the differences between non-interactive and interactive settings in stochastic convex optimization with local privacy. Learn about prior work on private stochastic convex optimization, including upper bounds and exponential sample complexity. Discover the lower bounds for non-interactive private linear classification and the statistical query model. Investigate the proof techniques used to demonstrate that learning with non-interactive local privacy requires exponential sample complexity. Gain insights into two non-distinguishable distributions over {-1,1}d and consider open questions in the field of distributed learning with privacy constraints.

Syllabus

Interaction is necessary for distributed learning with privacy or communication constraints Yuval Dagan, MIT Vitaly Feldman, Apple Research was done while at Google Brain
Non interactive vs. interactive
Learning with local privacy requires interaction
Learning with communication constraints
Outline
Prior work on private stochastic convex optimization Interactive setting . Non interactive . Upper bound, exponential sample complexity
Lower bound: stochastic convex optimization of linear models
Prior work on locally private linear classification Sample bounds for non-interactive algorithms Daniely, Feldman 18
Lower bound: non-interactive private linear classification Setting: non-interactive ; locally private with parameter e = 1 Thm: The sample complexity is exp min(d, 1/7)(1)
Proof Techniques
The statistical query model (SQ) [Kearns 93] . Instead of independent samples, can ask for statistics about the data . For example: mean, moments, expected loss, gradient
Non interactive statistical queries
Two non-distinguishable distributions over {-1,1}d
Summary and open questions • Learning with non-interactive local privacy requires exponential sample complexity

Taught by

Association for Computing Machinery (ACM)

Reviews

Start your review of Interaction Is Necessary for Distributed Learning with Privacy or Communication Constraints

Never Stop Learning.

Get personalized course recommendations, track subjects and courses with reminders, and more.

Someone learning on their laptop while sitting on the floor.