228499b55310264a8ea0e27b6e7c6ab6-AuthorFeedback.pdf

Wewillrevisethe9 sentencestomakeitclear.10 2. Comparison to polar andLDPC codes: We will add more figures for comparisons in the revised version. While TurboAE doesnotoutperform traditional16 codes, achieving a comparable reliability underdiscretechannels requires a major breakthrough. This is because17 learning an autoencoder including anon-differentiable layer (binarization of code) makes training challenging. Wesee35 (relatively) less examples at the decision boundaries, making it hard to train an accurate decoder.