Decentralized EM Algorithm for Gaussian Mixtures under Data Heterogeneity and Partial Labeling
Researchers have developed new decentralized algorithms for Gaussian mixture models in federated learning settings. These methods, including a momentum-based approach (MNEM) and a semi-supervised variant (semi-MNEM), address challenges posed by heterogeneous data distribution and partial labeling. Theoretical analysis suggests MNEM can achieve asymptotic efficiency comparable to centralized methods, while semi-MNEM enhances convergence speed, as demonstrated through simulations and an analysis of a chest X-ray dataset. AI
IMPACT Introduces novel algorithmic approaches for decentralized machine learning, potentially improving model accuracy and efficiency in distributed data scenarios.