Federated System and Scale Challenges - Google Meet 

Breakout session doc 

Moderators: Daniel Ramage and Mosharaf Chowdhury

How could federated technologies scale to a majority of the world’s computing tasks on private data? What are the emerging systems challenges that could prevent from scaling to billions of devices? Recognizing that privacy decisions are complex multi-faceted decisions, where would you typically set the bar in terms of privacy, if any, for federated solutions to use your data? With explicit consent? Without explicit consent? How could regulatory frameworks capture differences between on-device data use through federated technologies and server-side data use?

Federated Analytics - Google Meet 

Breakout session doc

Moderators: Stefano Mazzocchi and Ayfer Ozgur

What would it take to have SQL-like database computations on massively distributed datasets via a federated analytics framework? What are some canonical algorithms that we ideally want to have? What are the privacy challenges in this setting, and what is the role of interactivity? What key multi-party computation and trusted execution environment frameworks are important to strengthen the privacy and security of FA systems?

Federated Optimization - Google Meet 

Breakout session doc

Moderators: Brendan McMahan and Suhas Diggavi

How can we further push the frontiers of privacy/utility tradeoffs when training with differential privacy? Does the federated optimization problem change if the goal is to eventually give each device a personalized model? What theoretical characterizations of non-IID data across clients best capture real-world problems and predict real-world algorithm performance? Can FL benefit from recent progress on learned optimization algorithms?

Fairness & Bias in Federated Systems - Google Meet 

Breakout session doc

Moderators: Virginia Smith and Shanshan Wu

What are the main sources of unfairness and bias in production federated systems? Should unfairness be addressed algorithmically and via improved system designs? What are the relevant notions of fairness that should be used in federated learning? What does fairness mean in analytics applications? How can personalization be effectively used to reduce unfairness and improve the trained models for all users? How to address client sampling bias in large-scale federated systems?

Open Federated Experimentation Platforms - Google Meet 

Breakout session doc

Moderators: Zachary Garrett and Andrew Trask

Should there be many federated experimentation platforms, each specialized for certain applications and federated domains, or is it better to have one rich platform? What is the role of standardization? What are the main challenges in building scalable platforms for federated experimentation? How to allow researchers with limited compute resources to run experiments on largescale datasets? 

Federated Self-Supervised Learning - Google Meet 

Breakout session doc

Moderators: Raviteja Vemulapalli and Hang Qi

Recently, self-supervised learning approaches that train a generic feature extraction model using unlabeled data, have received significant attention from the ML community in the centralized setting. In this discussion session, we will focus on self-supervised learning in a federated setting. Which of the recent self-supervised learning approaches are suitable for a federated setting in which the amount of data on individual clients is limited and non-iid? Which approaches are more (computation and communication) resource friendly? How can we handle the inconsistencies between the representation spaces learned by individual clients in a round? What are the privacy implications of training a task-agnostic representation extractor in a federated setting?

Federated Technologies for IoT & Healthcare Applications - Google Meet 

Breakout session doc

Moderators: Stefan Mellem and Walid Saad

What would a decentralized world of computing, where the majority of data is processed and stored outside datacenters, look like? What is the role of sensors, mobile and stationary end-user devices, edge resources, and datacenters? What are the unique technical requirements and challenges associated with healthcare applications? What are practical challenges in deploying federated technologies for health research? When is federation the right tool for the job?

Trusted, Secure, and Robust Federated Aggregations - Google Meet 

Breakout session doc

Moderators: Adria Gascon, Kallista Bonawitz, Sewoong Oh

How do we achieve robust high-throughput learning under data heterogeneities while ensuring accuracy? What kinds of attacks on Federated Learning are actually practical and matter? Where are the sweet spots for secure computation (and cryptography in particular) in federated learning and analytics? Are there key functionalities, such as shuffling or (thresholded) aggregation that we should be focusing on? What are the requirements for the next generation of cryptographic protocols to enhance the privacy and security guarantees of FL. What will be the role of technologies such as trusted execution environments or zero knowledge systems?