We revisit the fundamental problem of physical layer communications, namely reproducing at one point a message selected at another point, to finally arrive at a trainable system that inherently learns to communicate and adapts to any channel environment. As such, we realize a data-driven system design, based on deep learning algorithms, leading to a universal framework that allows end-to-end optimization of the whole data-link without the need for prior mathematical modeling and analysis. A trainable communication system inherently tolerates and even exploits effects which are difficult to model, such as hardware imperfections and channel uncertainties. We show that such systems not only enjoy a competitive and in some cases even superior performance, but facilitate a simplified design flow due to their conceptual elegance and, hence, may trigger a paradigm shift of how we design future communication systems. We thus pose the seemingly simple, naive, yet in fact rather complicated and attractive research question: Can we learn to communicate?
The goal of this talk is to provide an introduction to the rapidly growing field of end-to-end learning of physical layer communications. For this, we reinterpret transceiver signal-processing blocks (e.g., quantization, coding, modulation, detection) as neural networks and show that this idea enables data-driven communication systems that perpetually learn and adapt to (m)any environment(s). Further, we show that the practical realization of end-to-end training of communication systems is fundamentally limited by its accessibility of the channel gradient. To overcome this major burden, the idea of generative adversarial networks (GANs) that learn to mimic the actual channel behavior has been recently proposed in the literature. Contrarily to handcrafted classical channel modeling, which can never fully capture the real world, GANs promise, in principle, the ability to learn any physical impairment, enabled by the data-driven learning algorithm. We verify the concept of GAN-based autoencoder training in actual over-the air (OTA) measurements.
In the second part of this talk, we show how training of autoencoder-based communication systems on the bit-wise mutual information allows seamless integration with practical bit metric decoding receivers, as well as joint optimization of constellation shaping and labeling. Additionally, we present a fully differentiable neural iterative demapping and decoding structure which achieves significant gains on additive white Gaussian noise channels. Going one step further, we show that careful code design can lead to further performance improvements.
Access: Seminar was delivered live via Zoom on October 2; watch the recording here.