Workshop on Decentralized Machine Learning, Optimization and Privacy (Sep 11-12, 2017)


With the advent of personal devices with computation and storage capabilities, it becomes possible to run machine learning on-device to provide personalized services to users without exposing their sensitive data to large data centers. Such decentralized architectures allow individuals to better control their data (with potential incentives for its usage), as well as reduce the infrastructure costs and risks related to data storage/processing for the service provider. This motivates the design of machine learning and optimization algorithms adapted to constraints arising from this new paradigm. Well beyond standard parallel computing techniques, it requires efficient solutions to deal with complex settings involving a very large number of parties, limited control over the network dynamics, heterogeneous local data distributions and/or the absence of a central coordinating entity. Another important challenge is to develop decentralized learning protocols which provably preserve privacy for each user and show some robustness against malicious parties.

This multidisciplinary workshop will be devoted to the new crucial scientific challenges raised by decentralized machine learning, including:

  • How to design efficient optimization algorithms (in terms of convergence rate, number of rounds, bandwidth, energy…) for the decentralized setting?
  • How can users collaborate to learn useful models in a fully decentralized network where communication is peer-to-peer only (no central entity)?
  • How to address privacy and security issues under various adversary models?

A major objective of the workshop is to initiate new fruitful collaborations between researchers in optimization, machine learning, privacy and distributed systems. Attendees are welcome to bring a poster to present their recent work at the poster session.

Location & Dates

The workshop will take place on September 11-12, 2017 at INRIA Lille (France). Lille is the capital of the north of France, a metropolis with 1 million inhabitants, with excellent train connection to Brussels (34 min), Paris (60 min) and London (80 min). See how to get there.


Registration is free but mandatory. Click here to register, preferably before July 15 (we may not be able to guarantee attendance for late registrations). Registration is now closed.

Confirmed speakers


Note: All events are located in the Amphitheater (Building B) except for the buffet lunch and poster session which are located in the Plenary room (Building A, across the street).

Monday, September 11th, 2017

09:30am – 10:00am Welcome (coffee and pastries)
10:00am – 12:00pm

Talk session (chair: Sébastien Gambs)

  • Stephen Hardy: Learn only the model [abstract] [slides]
  • Borja Balle: Privacy-Preserving Distributed Linear Regression on High-Dimensional Data [abstract] [slides]
12:00pm – 12:30pm Poster spotlights [poster list]
12:30pm – 02:30pm Buffet lunch + Poster session [poster list] (Plenary room, Building A across the street)
02:30pm – 04:30pm

Talk session (chair: Joseph Salmon)

  • Mikael Johansson: Sparsity and asynchrony in distributed optimization: models and convergence results [abstract] [slides]
  • Peter Richtárik: Privacy preserving randomized gossip algorithms [abstract] [slides]
04:30pm – 05:00pm Coffee break
05:00pm – 06:00pm

Talk session (chair: Morten Dahl)

  • Meilof Veeningen: Distributed Privacy-Preserving Data Mining in the Medical Domain [abstract]

Tuesday, September 12th, 2017

08:30am – 09:00am Welcome (coffee and pastries)
09:00am – 11:00am

Talk session (chair: Aurélien Bellet)

  • Keith Bonawitz: Federated Learning: Privacy-Preserving Collaborative Machine Learning without Centralized Training Data [abstract]
  • Dan Alistarh: Quantized Stochastic Gradient Descent [abstract] [slides]
11:00am – 11:30am Coffee break
11:30am – 12:30pm

Talk session (chair: George Giakkoupis)

  • Hamed Haddadi: User-Centric Personal Data Analytics on the Edge [abstract] [slides]


Financial support

The workshop is organized in the context of the ANR project PAMELA (Personalized and decentrAlized MachinE Learning under constrAints). We are grateful to our financial sponsors, listed below.

INRIA    Télécom ParisTech    ANR
   CRIStAL   CNRS Madics