MLWiNS: Decentralized Heterogeneous Deep Learning for Efficient Wireless Spectrum Monitoring

随着无线网络的日益庞大和复杂，传统的基于模型的信号处理技术的频谱监测方法已经变得不充分，甚至可能无法提供准确的无线网络评估。与此同时，深度学习技术在标准的集中式学习任务（如图像分类）中已经被证明是成功的，但对于大规模无线传感系统，它几乎没有被探索，这需要非常规的节点分布，复杂的信道衰落和用户协作机会。该项目为大规模无线系统开发创新的分散式异构深度学习技术。该项目的成果导致了技术创新，解决了最先进的无线传感和管理系统的几个主要挑战，包括传统传感和管理方案在超宽无线频谱设置中的无能性，使用深度学习技术处理异构任务和非iid数据的困难，以及大规模网络分布式深度学习中昂贵的通信和计算开销。该项目通过开发革命性的分散深度学习框架来解决大规模无线频谱传感的独特挑战。计划有三个主要推力。在推力1中，研究了当今复杂和大规模无线频谱传感的主要挑战，并为实际的频谱传感任务开发了一种基于深度学习的创新解决方案。在Thrust 2中，开发了专用的通信和计算方案来优化所提出的分散深度学习框架的性能。在Thrust 3中，基于对机器学习消耗无线网络资源为网络资源利用带来附加价值的关键观察，首次进行了探索性努力，以理解和利用机器学习在频谱管理中的复杂作用。实验测试证明了实际的频谱监测应用。所提出的无线传感和管理系统可以使大量的大型无线网络系统受益，例如5G无线网络和其他大型网状网络系统。该教育计划通过将嵌入式系统、移动计算和机器学习的跨学科模块与新开发的教学实践相结合，加强现有的课程和教学法。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

As wireless networks evolve to be increasingly massive and complex, traditional spectrum monitoring methods with model-based signal processing techniques have become inadequate and may even fail to provide accurate wireless network evaluation. Meanwhile, deep learning techniques have been proven successful in standard centralized learning tasks (e.g., image classification), yet it is barely explored for large-scale wireless sensing systems, which entail unconventional node distribution, complex channel fading and user collaboration opportunities. This project develops innovative decentralized heterogeneous deep learning techniques for large-scale wireless systems. The outcomes of this project lead to technical innovations that tackle several major challenges of the state-of-the-art wireless sensing and management systems, including the incapability of conventional sensing and management schemes in ultra-wide wireless spectrum settings, the difficulty in handling heterogeneous tasks and non-IID data with deep learning technologies, as well as the costly overhead of communication and computation in distributed deep learning for large-scale networks.This project addresses the unique challenges of large-scale wireless spectrum sensing by developing a revolutionary decentralized deep learning framework. Three main thrusts are planned. In Thrust 1, major challenges of complex and large-scale wireless spectrum sensing nowadays are investigated, and an innovative deep learning-based solution is developed for practical spectrum sensing tasks. In Thrust 2, dedicated communication and computation schemes are developed to optimize the performance of the proposed decentralized deep learning framework. In Thrust 3, the very first exploratory effort is made to understand and utilize the intricate role of machine learning in spectrum management, based on the key observation that it consumes wireless network resources to bring in added value to network resource utilization. Experimental testing is demonstrated for practical spectrum monitoring applications. The proposed wireless sensing and management system can benefit a plethora of large-scale wireless network systems, such as a 5G wireless network and other large-scale mesh networking systems. The education plan enhances existing curricula and pedagogy by integrating interdisciplinary modules on embedded systems, mobile computing, and machine learning with newly developed teaching practices.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.