Computationally-Efficient Sparsely-Connected Multi-Output Neural Networks for IM/DD System Equalization

Low-complexity sparsely-connected multi-output neural networks are proposed for equalization in a 50-Gb/s 25-km PAM4 IM/DD system. Compared with traditional fully-connected single-output counterparts, a gross complexity reduction of 60.4%/56.7% can be achieved with 2-layer FNN/C-FNN architecture. © 2022 The Author(s)