Dr. Christian Häger (December 13, 2017 at 1:00 PM, LNT Library N2405)
On December 13, 2017 at 1:00 PM, Dr. Christian Häger from Department of Electrical Engineering at Chalmers University of Technology, Sweden will be giving a talk in the LNT Library N2405 about "Coding and Deep Learning for High-Speed Fiber-Optic Communication Systems".
Coding and Deep Learning for High-Speed Fiber-Optic Communication Systems
Dr. Christian Häger
Chalmers University of Technology, Sweden
Fiber-optic communication links carry virtually all intercontinental data traffic and are often referred to as the Internet backbone. The first part of this talk is about staircase codes which are error-correcting codes tailored to the specific requirements of fiber-optic systems. In particular, they achieve very high throughputs (100 Gb/s), provide large net coding gains, and have extremely low error floors. We show that, despite their deterministic structure, the asymptotic performance of staircase codes under iterative decoding can be rigorously analyzed using density evolution. This result can be used, for example, to rapidly predict the performance or efficiently optimize code parameters. We also show that the performance of staircase codes can be improved by using a novel anchor-based decoder which addresses the problem of miscorrections on the binary symmetric channel.
The goal of the second part is to illustrate how deep learning can be used to equalize linear and nonlinear channel impairments in fiber-optic systems. In particular, the optical waveform channel is defined by the nonlinear Schrödinger equation. Channel equalization can be performed by numerically solving this equation using the received signal as a boundary condition. This approach is commonly referred to as digital backpropagation. Interestingly, applying the popular split-step Fourier method for this task leads to a computation graph that is mathematically equivalent to a deep neural network. This observation allows us to jointly optimize all linear filtering steps using deep learning. The resulting "learned" digital backpropagation can reduce the complexity by over an order of magnitude compared to conventional implementations of the split-step Fourier method.
Based on joint work with Henry D. Pfister, Alexandre Graell i Amat, Fredrik Brännström, and Erik Agrell. Preprints available at:
Christian Häger received the Dipl.-Ing. degree in electrical engineering from Ulm University, Germany, in 2011 and the Ph.D. degree in communication engineering from Chalmers University of Technology, Sweden, in 2016. Since August 2016, he is a postdoctoral researcher at the Department of Electrical and Computer Engineering at Duke University, USA, and, since April 2017, also at the Department of Electrical Engineering at Chalmers University of Technology. Currently, he works with Prof. Henry D. Pfister (Duke University) and Prof. Alexandre Graell i Amat (Chalmers University) on modern coding theory and machine learning for high-speed fiber-optic communication systems. In 2017, he received the Marie Skłodowska-Curie Individual Fellowship from the European Commission.