Communication systems using neural network-based transceivers with autoencoder-driven end-to-end learning

在过去的几年里，机器学习领域，特别是深度学习领域有了非常迅速的发展；它们的应用现在扩展到几乎每个行业和研究领域。尽管研究人员过去曾试图用机器学习来解决与通信相关的问题，但它仍然没有对我们今天设计和实现通信系统的方式产生根本影响。乍一看，机器学习技术似乎并不能很好地与物理层上的通信相匹配，50年来，机器学习技术在信号处理、通信和信息理论的基础上取得了巨大的进步，在许多渠道上接近最优的香农极限性能。然而，仍然有几个悬而未决的问题，例如关于联合处理的适应性和复杂性，其中使用基于机器学习的方法的第一个结果是有希望的。该提案旨在审查基于“自动编码器”概念和深度学习技术的可学习的端到端通信。它承诺了一种通信系统，它可以学习在任何类型的信道上进行通信，而不需要事先对信道模型进行详细的数学抽象，通过从手工制作的、仔细优化的子块转向自适应和灵活的(人工)神经网络，打破了传统基于块的信号处理中常见的限制，导致了许多有吸引力的研究问题。为了更全面地了解机器学习技术在通信中的潜力，我们从经典的信号处理作为参考；然后我们使用传统的基于块的学习(例如替换经典的调制、检测或均衡块等)来研究神经网络，直到最终得到基于自动编码器驱动的端到端学习的多块神经网络。我们还计划通过空中测量来验证所探索的概念，从而在通信渠道上产生许多在许多经典模型中找不到的影响，因此需要隐式学习。机器学习方法的好处可能包括更灵活的硬件、高度适应性的系统和更低的总体复杂性。因此，我们提出了一个看似幼稚但实际上相当复杂和吸引人的研究问题：“我们能学会通信吗？”请注意，这个建议针对的是物理层传输，而没有进一步考虑消息本身的语义(即，没有训练对消息或其内容的“理解”)。

The fields of machine learning and, in particular, deep learning have seen very rapid growth during the past few years; their applications now extend into almost every industry and research domain. Although researchers have tried to address communications-related problems with machine learning in the past, it still has had no fundamental impact on the way we design and implement communications systems today. At first glance, machine learning techniques do not appear to be a good match to communications on the physical layer, with 50 years of tremendous progress based on "classic" signal processing, communication and information theory, approaching close-to-optimal Shannon limit performance on many channels. However, several open problems remain, e.g. pertaining adaptivity and complexity of joint processing, where first results using machine learning-based approaches are promising. This proposal seeks to examine learnable end-to-end communications based on the "autoencoder" concept and deep learning techniques. It promises a communications system that can learn to communicate over any type of channel without the need for detailed prior mathematical abstraction of the channel model, breaking up restrictions commonplace in conventional block-based signal processing by moving away from handcrafted, carefully optimized sub-blocks towards adaptive and flexible (artificial) neural networks, leading to many attractive research questions. To obtain a more comprehensive understanding of the potential of machine learning techniques for communications, we start off from classic signal processing as a reference; then we study neural networks using conventional block-based learning (replacing, e.g. classic modulation, detection, or equalization blocks, ...), until finally arriving at multi-block neural networks based on autoencoder-driven end-to-end learning. We also plan to validate the explored concepts by over-the-air measurements, giving rise to many effects on the communication channel that cannot be found in many classical models, and, thus, need to be learned implicitly. The benefits of machine learning approaches may include more flexible hardware, highly adaptive systems and less overall complexity. We thus pose the seemingly naive, yet, in fact, rather complicated and attractive research question: "Can we learn to communicate?"Note that this proposal targets physical layer transmission without any further respect to the semantics of the message itself (i.e. no “understanding” of the message or its content is trained).