Abstract

End-to-end learning is a promising solution to realize the optimal performance of optical communication systems. By replacing the independent signal processing modules in the transmitter and receiver with deep neural networks, end-to-end system optimization can be achieved via training the neural networks together on a differentiable channel. In this paper, a noise adaptation network for channel modeling is proposed to simulate channel response and the impact of channel noise on transmitted signals. The structure of the noise adaptation network is multi-scale deep neural network (MscaleDNN), which can better characterize the channel at different frequencies. Based on the noise adaptation network, a novel end-to-end learning framework is further designed. Within the framework, memory buffer technology and constraint loss are introduced to significantly enhance the efficiency and performance of end-to-end learning. Experimental demonstration of the proposed end-to-end learning scheme is performed on a 100G passive optical network (PON) system based on intensity modulation and direct detection. The results indicate that, compared to the optimized Volterra non-linear equalization at the receiver and the joint equalization achieved by indirect approach, end-to-end optimization improves receiver sensitivity by 0.8 dB and 1.2 dB, respectively, and achieves a power budget of 31.4 dB. In particular, the advantages of the end-to-end learning are even more pronounced in the case of higher received optical power.