Neural Augmented Min-Sum Decoding of Short Block Codes for Fading Channels

In the decoding of linear block codes, it was shown that noticeable gains in terms of bit error rate can be achieved by introducing learnable parameters to the Belief Propagation (BP) decoder. Despite the success of these methods, there are two key open problems. The first is the lack of analysis for channels other than AWGN. The second is the interpretation of the weights learned and their effect on the reliability of the BP decoder. In this work, we aim to bridge this gap by looking at non-AWGN channels such as Extended Typical Urban (ETU) channel. We study the effect of entangling the weights and how the performance holds across different channel settings for the min-sum version of BP decoder. We show that while entanglement has little degradation in the AWGN channel, a significant loss is observed in more complex channels. We also provide insights into the weights learned and their connection to the structure of the underlying code. Finally, we evaluate our algorithm on the over-the-air channels using Software Defined Radios.