Non-asymptotic Evaluation of Federated EM Algorithms


Obtain a PDF of the paper titled In direction of the Idea of Unsupervised Federated Studying: Non-asymptotic Evaluation of Federated EM Algorithms, by Ye Tian and a couple of different authors

Obtain PDF

Summary:Whereas supervised federated studying approaches have loved important success, the area of unsupervised federated studying stays comparatively underexplored. A number of federated EM algorithms have gained reputation in apply, nonetheless, their theoretical foundations are sometimes missing. On this paper, we first introduce a federated gradient EM algorithm (FedGrEM) designed for the unsupervised studying of combination fashions, which dietary supplements the present federated EM algorithms by contemplating process heterogeneity and potential adversarial assaults. We current a complete finite-sample concept that holds for normal combination fashions, then apply this normal concept on particular statistical fashions to characterize the specific estimation error of mannequin parameters and combination proportions. Our concept elucidates when and the way FedGrEM outperforms native single-task studying with insights extending to current federated EM algorithms. This bridges the hole between their sensible success and theoretical understanding. Our simulation outcomes validate our concept, and reveal FedGrEM’s superiority over current unsupervised federated studying benchmarks.

Submission historical past

From: Ye Tian [view email]
Mon, 23 Oct 2023 19:53:36 UTC (1,157 KB)
Mon, 5 Feb 2024 05:39:28 UTC (523 KB)

Supply hyperlink


Please enter your comment!
Please enter your name here