Extrinsic Neural Network Equalizer for Channels with High Inter-Symbol-Interference

In this paper, we propose a novel extrinsic neural network equalizer (ExNE) for joint iterative equalization and decoding. The proposed ExNE takes the received signal sequence and a priori probabilities from the channel decoder as inputs to directly generate output extrinsic probabilities. This approach improves the performance of iterative equalization and decoding by making explicit use of extrinsic information. A three-step, open-loop neural network (NN) training procedure is developed for the ExNE, independent of the choice of channel code. We propose a new NN architecture termed deep concatenated convolutional blocks with skip connections (DCCB-SC) for ExNE which attains excellent performance with only a moderate number of network parameters. For challenging linear and non-linear inter-symbol-interference (ISI) channels considered in this work, the proposed ExNE approaches the performance of the maximum a posteriori probability (MAP) equalizer without assuming prior knowledge of the channel model.