Sebastian Dörner

Deep Learning for Communications

The huge success of deep learning (DL) and neural networks (NNs), mainly in the fields of computer vision and speech processing, has recently triggered further exploration of DL for communications. The potential applications span from trainable channel decoders, NN-based multiple-input multiple-output (MIMO) detectors and user position estimation based on channel state information up to communication systems that inherently learn to communicate, so called autoencoder communication systems.

Since the first publication of autoencoder-based communications by O'Shea and Hoydis, the vision of end-to-end training of communication systems has attracted an impressive amount of follow-up work. Thereby, end-to-end learning has found its entry into virtually any field of today’s communications research – in the wireless and optical domain, but also emerging domains, like the molecular channel. Although autoencoder-based communication promises a framework that operates over any channel, for practical deployment the missing channel gradient prevents joint end-to-end training of transmitter and receiver. To overcome this major obstacle of end-to-end training, several approaches (ranging from model based to GAN-based based training schemes) have been proposed and investigated at our institute.

We have shown the practicability and universality of the autoencoder framework in several publicatoins, including the demonstration of the world’s first communication system that solely consists of NN-based signal processing at Bell Labs Future X Days in Paris 2017.

The main goal of my research is the combination of classical algorithms in communications with trainable components and deep learning techniques, to improve state of the are baselines (neural network augmented algorithms), wherever possible. While classical solutions are often optimal within their expected operational scope, the idea is to improve them with NNs to either achieve a wider range of operation or combine certain processing steps that are classically done separately with a known performance loss. With the autoencoder framework being the most radical approach, as is aims on replacing most classical signal processing in transmitter and receiver by a single NN.