Abstract: Decentralized Multi-agent Learning (DML) enables collaborative model training while preserving data privacy. How-ever, inherent heterogeneity in agents' resources (computation, communication ...