Learning to Communicate: Deep Learning based solutions for the Physical Layer of Machine Type Communications [LeanCom]

随着物联网（IoT）、机器类型通信（MTC）、云计算和许多其他应用的出现，无线网络将变得更加复杂，同时也比以往任何时候都更加重要。 鉴于无线系统的连接性和复杂性的上述指数增长以及对延迟、容量、超可靠性和安全性的前所未有的需求，网络在分析上变得难以处理。自然地，基于驱动当今网络架构的通信系统和网络的数学模型的人为驱动的物理层（PHY）设计方法正被新兴网络范例的纯粹复杂性所超越。使用低成本MTC传感器和发射器不可避免的硬件缺陷将大大增加网络的波动性，并且通常依赖于通用和高度不准确模型的理论驱动的解决方案无法解决这个问题，因为它们在实践中是高度次优的。上述挑战需要新的数据驱动的方法来设计通信系统，而不是传统的系统模型驱动的设计，正变得过时。朝着未来的MTC的多样化的通信范例，迫切需要解决可靠的和自适应的链接脱离数学模型，而不是基于数据驱动的方法。这个富有远见的项目将通过开发为无线通信量身定制的新神经网络架构和基于数据驱动训练的新收发器架构来解决这些基本挑战。我们的研究将致力于a）通信特定DL框架的开发，B）DL启发的PHY解决方案，以及c）所提出的解决方案的概念验证。LeanCom将与华为，NEC欧洲，杜克大学，数字弹射器和CommNet，并渴望启动一个创新的生态系统，（c）鼓励信通技术基础设施和服务提供者中的有影响力的参与者开发新一代学习型网络并使之商业化。拟议解决方案的实施，实验和测试（在WP3内）作为LeanCom成果商业化的平台，旨在对英国数字经济的基础性影响。

With the advent of the Internet of Things (IoT), machine type communications (MTC), cloud computing and many other applications, the wireless network will become far more complex, while at the same time far more essential than ever before. Given the above exponential growth in both connectivity and complexity of the wireless systems and the unprecedented demands on latency, capacity, ultra-reliability and security, the network is becoming analytically intractable. Naturally, human-driven physical layer (PHY) design approaches rooted on mathematical models of communications systems and networks which drive today's network architectures are being surmounted by the sheer complexity of the emerging network paradigms. Hardware imperfections, that are inevitable with the employment of low-cost MTC sensors and transmitters, will drastically increase the volatility of the network, and theoretically driven solutions typically relying on generic and highly inaccurate models cannot address this as they are highly sub-optimal in practice. The above challenges necessitate new data-driven approaches to the design of communications systems, as opposed to traditional system-model driven designs that are becoming obsolete.Towards the diverse communication paradigms of MTC of the future, there is an urgent need to address reliable and adaptive links detached from mathematical models, and instead based on data-driven approaches. This visionary project will address these fundamental challenges by developing new Neural Netowrk architectures tailored for wireless communications, and new transceiver architectures based on data-driven training. Our research will address the development of a) a communications specific DL framework, b) DL-inspired PHY solutions and, c) proof-of-concept verification of the proposed solutions.LeanCom will be performed with Huawei, NEC Europe, Duke University, The Digital Catapult and CommNet and aspires to kick-start an innovative ecosystem for high-impact players among the infrastructure and service providers of ICT to develop and commercialize a new generation of learning-based networks. The implementation, experimentation and testing (within WP3) of the proposed solutions serves as a platform towards commercialisation of the results of LeanCom, aiming towards an impact of a foundational nature for the UK's digital economy.