Parallelization of Recurrent Neural Network-Based Equalizer for Coherent Optical Systems via Knowledge Distillation
Sasipim Srivallapanondh, Pedro J. Freire, Bernhard Spinnler, Nelson Costa, Antonio Napoli, Sergei K. Turitsyn, Jaroslaw E. Prilepsky
Abstract
The recurrent neural network (RNN)-based equalizers, especially the bidirectional long-short-term memory (biLSTM) structure, have already been proven to outperform the feed-forward NNs in nonlinear mitigation in coherent optical systems. However, the recurrent connections still prevent the computation from being fully parallelizable. To circumvent the non-parallelizability of recurrent-based equalizers, we propose, for the first time, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">knowledge distillation</i> (KD) to recast the biLSTM into a parallelizable feed-forward 1D-convolutional NN structure. In this work, we applied KD to the cross-architecture regression problem, which is still in its infancy. We highlight how the KD helps the student's learning from the teacher in the regression problem. Additionally, we provide a comparative study of the performance of the NN-based equalizers for both the teacher and the students with different NN architectures. The performance comparison was carried out in terms of the Q-factor, inference speed, and computational complexity. The equalization performance was evaluated using both simulated and experimental data. The 1D-CNN outperformed other NN types as a student model with respect to the Q-factor. The proposed 1D-CNN showed a significant reduction in the inference time compared to the biLSTM while maintaining comparable performance in the experimental data and experiencing only a slight degradation in the Q-factor in the simulated data.