DRACO: Decentralized Asynchronous Federated Learning over Continuous Row-Stochastic Network Matrices

Recent advancements in machine learning, networked intelligent systems, and wireless connectivity have paved the way for various innovative applications and use cases across various sectors, including the Internet of Things (IoT), consumer robotics, autonomous transportation, and edge computing. These systems increasingly rely on decentralized learning architectures for processing data where generated, minimizing latency and bandwidth usage while enhancing privacy . However, these benefits come with significant challenges, particularly in terms of ensuring efficient and reliable communication and processing within inherently unstable and diverse network environments. Addressing these challenges requires novel approaches that adapt to the unique demands of decentralized architectures, fostering robust and expandable solutions for real-time data processing and learning. In this work, we consider the problem of communication efficiency in federated learning (FL) [1] and in particular in serverless (fully decentralized) learning settings that operate without a central coordinating server [2-6]. Asynchronous learning, empowering each participant to conduct local training and data transmission at their own pace, is a standard and relevant design choice in decentralized network schemes [7-12]. Asynchronous and decentralized learning have an advantage when used separately from each other, manifesting as adaptability to limited resources and downsized communication overhead. Y et unfortunately, when these two paradigms are combined, their integration poses a greater challenge in achieving a unanimous global consensus, as required for instance in the development of sophisticated navigation algorithms [13]. Decentralized optimization studies in the literature often involve high "synchronization costs" due to the complexity of ensuring consensus.