Key Enablers for Hardware-Efficient Neural Network Receiver in Short-Reach Optical Links

This paper systematically studies hardware-efficient neural network-based equalization techniques for short-reach direct-detection optical systems. Transfer learning can be utilized for accelerating the training procedure, while pruning, multi-task learning and quantization are proposed for efficient equalization.