Neural Information Bottleneck Decoding

Abstract

Receiver-sided channel decoding is a crucial, but computationally very demanding task. Recently, information-bottleneck-based decoding received considerable attention in the literature, as it achieves very good performance with coarse quantization and low complexity. Information bottleneck de-coders replace the conventional node operations in the channel decoders with relevant-information-preserving mappings. In the literature, these mappings are typically represented as lookup tables. Unfortunately, the lookup tables are fairly large and the logic circuits needed to implement the lookup tables occupy significant chip area. This paper proposes a novel approach to overcome these impairments, which is based on a trainable function, i.e., a neural network. We show that a trained neural network can represent the mappings more compactly than the frequently used lookup tables. Moreover, our results show that even multiple-input-multiple-output mappings can be learned to replace lookup tables with prohibitive complexity. We investigate a coarsely quantized low-density parity-check decoder that realizes the relevant-information-preserving mappings for all decoding iterations with a single neural network to illustrate the practical benefits of the proposed concept.