Local Epochs as Optimal Stopping: Theory and Evidence in Federated Learning

Key points

• How long should each client train before communicating?1
• We cast local training as finite-horizon optimal stopping: a client trains while the next epoch's expected loss reduction exceeds its priced squared drift.1
• We prove that this one-stage look-ahead threshold is optimal whenever stopping decisions are nested, derive closed forms for geometric and power-law returns, and show that an estimated rule loses utility only through wrong-signed decisions.1
• Epoch-level evidence supports the model's structure (geometric diminishing returns and additive drift) and reveals a simple regularity: the stage of training, not the client, sets the budget.1

Sources (1)

• [1]Local Epochs as Optimal Stopping: Theory and Evidence in Federated Learning
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 15, 01:15 PM

  Federated learning usually answers with a fixed, hand-tuned number of local epochs, yet an extra epoch helps only while the loss it removes outweighs the client drift it adds.

  How long should each client train before communicating?

Extractive summary: sentences quoted from the sources.