Efficient Deep Learning of Nonlinear Fiber-Optic Communications Using a Convolutional Recurrent Neural Network

A hybrid architecture comprising a CNN encoder and a many-to-one RNN was proposed.

CNN captures short-temporal dependencies, and RNN extracts long-term features.

CNN lessens the computational burden on RNN by taking the data into a latent space.

Thanks to the shrunk number of time-steps, a many-to-one RNN works sufficiently.

In a 100 GBd 16-QAM 20 x 100 km SMF optical transmission system, the CNN+RNN resulted in ~98% FLOPs reduction over Learned DBP, and more than 50% FLOPs reduction over the state-of-the-art CNN+bi-LSTM approach.