Federated Semi-Supervised Learning with Uniform Random and Lattice-Based Client Sampling.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Distributed Systems

Background:

• Federated semi-supervised learning (Fed-SSL) utilizes distributed labeled and unlabeled data.
• Partial client participation is common in Fed-SSL to reduce communication overhead.
• Non-i.i.d. data distributions pose challenges for client sampling strategies in Fed-SSL.

Purpose of the Study:

• To propose a novel federated averaging semi-supervised learning algorithm, FedAvg-SSL.
• To investigate the impact of different client sampling strategies on Fed-SSL performance.
• To analyze the convergence properties of the proposed algorithm.

Main Methods:

• Introduced FedAvg-SSL, incorporating uniform random sampling (Monte Carlo) and lattice-based sampling (quasi-Monte Carlo).
• Clients alternate between updating the global model and refining a pseudo-label model using local data.
• Provided theoretical convergence analysis and conducted extensive experiments.

Main Results:

• FedAvg-SSL achieves a sublinear convergence rate with linear speedup.
• Lattice-based sampling demonstrates advantages over uniform random sampling in federated learning.
• Experimental results validate theoretical findings and highlight the impact of sampling strategies.

Conclusions:

• FedAvg-SSL offers an effective approach for federated semi-supervised learning.
• Lattice-based sampling enhances training stability and model performance, especially under non-i.i.d. conditions.
• The study provides insights into optimizing client participation and sampling for Fed-SSL.