Invited paper: Towards robust end-to-end neural network-based transceivers for short reach fiber links

Short reach fiber links based on the intensity modulation/direct detection (IM/DD) technology require novel digital signal processing (DSP) solutions for compensating severe impairments such as fiber dispersion and photodiode nonlinearity. Recently, end-to-end deep learning-based autoencoder systems where the complete transceiver is implemented as a single neural network (NN) were demonstrated, both numerically and on experimental test-beds, as a viable alternative to conventional DSP for the optical IM/DD links. In particular, it has been shown that carefully chosen autoencoder architectures have the potential to provide performance improvement as well as complexity reduction. Nevertheless, one of the challenges in designing NN-based transceivers for economical short reach fiber links lies in developing simple yet efficient offline optimization procedures which allow operation over various link conditions without requiring additional on-device learning, thus circumventing a substantial complexity overhead. This paper investigates numerically a low-complexity offline end-to-end learning strategy for achieving significant robustness to dispersion (distance) variations in NN-based transceivers for short reach. More specifically, the training procedure can enable transmission with bit error rates below a common hard-decision forward error correction threshold over a range of distances around the nominal link distance.