AsylADMM: A Novel Gossip Algorithm for Robust Decentralized Learning

Researchers have introduced AsylADMM, a new asynchronous gossip algorithm designed to enhance decentralized learning on resource-constrained edge devices. Current state-of-the-art methods struggle to balance communication efficiency with robustness against data corruption, particularly when handling non-smooth objectives like pinball or l1 loss. Existing solutions often require excessive memory that scales with node degree, rendering them impractical for limited hardware. AsylADMM addresses this by requiring only two variables per node, significantly reducing memory overhead while maintaining robustness. The study provides theoretical analysis for the synchronous variant and proves convergence in simplified settings. Empirical results demonstrate that AsylADMM converges faster than existing baselines on challenging tasks, including quantile estimation, geometric median estimation, Lasso regression, and robust regression. This development offers a practical pathway for implementing robust, non-smooth decentralized learning systems, making advanced machine learning more accessible and efficient for edge computing environments where memory and communication bandwidth are critical constraints.