On Robust Deep Neural Decoders

Meryem Benammar, Pablo Piantanida

Abstract: The design of neural-based decoders is well understood for classical point to point channels such as the Additive White Gaussian Noise (AWGN) channel, the Binary Symmetric Channel (BSC) and the Binary Erasure Channel (BEC). For such channels, an optimal training noise distribution allows the neural decoder to generalize to other channel parameters unseen during training, i.e., other values of Signal to Noise Ratios (SNR), crossover probabilities or erasure probabilities. However, we show in this work that, for other families of channels such as the Binary Asymmetric Channel (BAC), the training and validation mismatch can be detrimental to the performances of the learned decoder. We investigate more robust constructions of neural decoders based on Domain Adaptation (DA) techniques.