Optimal Training Channel Statistics for Neural-based Decoders

This work investigates the design of End-to-End channel coding based on deep learning. The focus is on the design of neural networks based channel decoders. We demonstrate the existence of an optimal training statistic for the cross-entropy loss which allows the network to generalize to channel statistics unseen during training while performing close to their optimal decision rule. Numerical results illustrate an application to Polar coding on binary input memoryless channels.