Reservoir Computing-based Multi-Symbol Equalization for PAM 4 Short-reach Transmission

The intensity-modulated directly detected (IM/DD) links are the primary candidates for such short-reach interconnects [1]. However, the IM/DD links with square-law photodetection (PD) convert chromatic dispersion (CD) into a nonlinear inter-symbol interference (ISI) accumulated during the propagation in fiber. To mitigate the ISI, recurrent neural networks (RNNs) with sliding window input have been proposed in [2], achieving substantial bit-error-rate (BER) performance. However, the RNNs incorporate high computational complexity (CC) in terms of real multiplications per equalized symbol (RMPS) and are challenging to train [3]. Therefore, reservoir computing (RC), a special type of RNN, has been introduced in [4,5] to compensate for fiber-induced impairments, while keeping computational and training complexity low. In [6], it was shown that introducing a sliding window at the input boosts the RC performance for time series prediction tasks. Usually, in NNs-based digital signal processing blocks, the equalization is performed sequentially symbol by symbol [7]. However, it was experimentally demonstrated in [3] that feedforward NN-based multi-symbol equalization achieved considerable CC relaxation without losing in BER. In this work, we inherit the idea of using a sliding window at the input of the RC and propose a multi-symbol RC equalizer to decrease CC without impairing performance.