Deep-Learning-Enabled Direct Detection With Reduced Computational Complexity and High Electrical-Spectral-Efficiency

Complex-valued double-sideband (CV-DSB) direct detection (DD) is a promising solution for high capacity and cost-sensitive data center interconnects, since it can reconstruct the optical field as a homodyne coherent receiver while does not require a costly local oscillator laser. In a carrier-assisted CV-DSB DD system, the carrier occupies a large proportion of the total optical signal power but bears no information. Reducing the carrier to signal power ratio (CSPR) can improve the information-bearing signal power and thus maximize the achievable system performance. Recently, we have proposed and demonstrated a deep-learning-enabled DD (DLEDD) scheme to reconstruct the full-field of the CV-DSB signal. In the DLEDD scheme, the optical CV-DSB signal was detected by a dispersion-diversity receiver and then recovered by a deep convolutional neural network (CNN). Nevertheless, the computational complexity of the deep CNN is the main obstacle to the application of the DLEDD scheme. In this paper, we demonstrate a 50-GBaud CV-DSB 32-ary quadrature amplitude modulation (32-QAM) signal transmission over 80-km single-mode fiber with similar to 64% computational-budget reduction in the field reconstruction. This is achieved by using 1x1 convolutions to attain a sparse dimensionality (in particular the number of channels) of the deep CNN. To the best of our knowledge, we achieve the highest electrical spectral efficiency of 7.07 b/s/Hz per polarization per wavelength for a CV-DSB DD receiver without requiring a sharp-roll-off optical filter.