Polarized Dropout: A Novel Deep Joint Source Channel Coding Scheme for Erasure Channels

This letter investigates the challenges encountered by deep joint source-channel coding in erasure channels. We explore the effectiveness of the widely adopted dropout technique in endowing deep neural networks with resilience against erasures. However, directly applying dropout at the channel layer introduces uncertainty into the neural network’s training process, leading to performance degradation. To address this issue, we introduce the Polarized Dropout scheme and a novel network architecture that encodes analog symbols using the Walsh-Hadamard transform based on real-number field computation. Leveraging the polarization of symbol recovery probabilities, for a given erasure rate, a determined set of neurons will be assigned a dropout rate of 1, while the remainder are assigned a dropout rate of 0. Simulation results indicate a maximum enhancement of nearly 6dB in communication performance.