Oort: Efficient Federated Learning via Guided Participant Selection

https://www.usenix.org/conference/osdi21/presentation/lai

Trends:

• Edge devices: massive data

• Increasing resource on edge devices

Federated Learning on the Edge

• On-device ML helps

Execution of FL

• Objective:

  • Better time to accuracy

• Step

  • Client selection

  • Execution

  • Aggregation

Challenge

• FL: heterogeneous

  • data distribution

• Existing work

  • System efficiency

    • Round duration

  • Statistical efficiency

    • number of rounds to accuracy

    • Determined by the client data

Problem

1. Overlook heter, client utility --> suboptimal training convergence

2. Unable for selection criteria --> useless testing results

Design

1. Enable faster FL training

   1. Challenge: identify heterogenous client utility, select high-utility clients at scale, select high-utility clients adaptively

2. Support interpretable FL testing

RL:

1. Action space / reward space: dynamics of the action space (might drop)

2. Training overhead