Polar Auto-Decoder With Parameterized Optimization Based on Structured End-to-End Bayesian Neural Network in IM/DD Optical Interconnection

This paper proposes an auto-decoder application employing end-to-end (E2E) learning in intensity-modulation direct-detection (IM/DD) optical interconnection system. Aiming to enhance the feedforward error correction (FEC) reliability with optimized decoding complexity, the polar auto-decoder is explored based on structured E2E Bayesian neural network (BNN) with hybrid model- and data-driven architecture, instead of only optimizing specific components of the communication system. Firstly, the scaling parameters of the model-driven BNN are incorporated into the belief propagation (BP) polar decoder by assigning weights to the layers of the factor graph. Secondly, through the E2E training considering the global impairments in fiber channel, including attenuation, dispersion, and nonlinear distortions, the data-driven components in the polar decoder are shared across different iterations, thus achieving reduced complexity due to the significantly optimized decoding efficiency and error correction performance. Experimental results within 80 GBaud PAM4 signal transmission over single-mode fiber reveal that the proposed auto-decoder employing structured E2E learning strategy obtains 0.7-dB improvement in received power sensitivity, while reducing memory overhead by 80% relative to the MS-DNN decoder and maintaining decoding performance.