On the Computational Complexity of Artificial Neural Networks for Short-Reach Optical Communication

Artificial neural networks (ANNs) have been widely used for nonlinear equalization in short-reach optical communications due to the superior performance compared with traditional approaches. However, the computational complexity (CC) is a major challenge that hinder their applications. In this paper, we compare various types of ANNs for short-reach optical communication with CC constraint, as well as provide several means of CC reduction by employing transfer learning, pruning and multi-task networks.