Asynchronous Parallel Incremental Block-Coordinate Descent for Decentralized Machine Learning

Machine learning (ML) is a key technique for big-data-driven modelling and analysis of massive Internet Things (IoT) based intelligent ubiquitous computing. For fast-increasing applications data amounts, distributed promising emerging paradigm since it often impractical or inefficient to share/aggregate centralized location from distinct ones. This paper studies the problem training an ML model over decentralized systems, where are many user devices algorithm run on-device, with aim relaxing burden at central entity/server. Although gossip-based approaches have been used this purpose in different use cases, they suffer high communication costs, especially when number large. To mitigate this, incremental-based methods proposed. We first introduce incremental block-coordinate descent (I-BCD) ML, which can reduce costs expense running time. accelerate convergence speed, asynchronous parallel BCD (API-BCD) method proposed, multiple devices/agents active fashion. derive properties proposed methods. Simulation results also show that our API-BCD outperforms state art terms time costs.