Year: 2026
Type: dissertation
Abstract: Recent advances in deep learning show that model performance consistently improves with larger and more diverse datasets.However, in many practical scenarios, a single data holder lacks sufficient data to benefit from such scaling.Collaborative learning, in which multiple parties jointly train a model using their combined data resources, offers a promising solution, yet direct data sharing is often infeasible due to privacy and confidentiality constraints.Federated learning provides a privacypreserving framework by enabling model training across distributed data without exchanging raw data.Despite its potential, collaboration remains fundamentally challenged by data heterogeneity, including differences in underlying data distributions across parties and inconsistencies in observable features or label availability, whether or not the data are centrally shared.Table 5.3: Ablation on cluster weights calculation.The FedWCA's cluster weights calculation is replaced by various methods such as one-hot weights (One-hot), equal weights (Equal), a hybrid of them (One-Equal), an adaptive version of One-Equal (One-Equal-A), and a local calculation-only FedWCA variant described in Section 5.4.2 (FedWCA-L).The original FedWCA surpasses all other methods across all PACS source domains. Methods
Author: Mori, Junki
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204240603