End-to-end learning strategy with channel-aided polar based autoencoder in IM/DD optical interconnection

An end-to-end (E2E) learning strategy with a channel-aided polar autoencoder (AE) is proposed and experimentally demonstrated in an intensity-modulation direct-detection optical interconnection system. With the global training of the E2E deep neural network, the learned fiber channel and transceiver characteristics are used to back-forward aid the channel indexes of AE, thus obtaining optimal polarization weight in polar encoding to mitigate signal impairments. The experimental results reveal that when the 50 GBaud PAM4 signals are transmitted over standard single mode fiber, the proposed polar AE by E2E learning strategy can effectively improve the received power sensitivity by 1.5 dB under conditions of overall adaptation of system parameters optimization.