Sparse Dense Neural Network Architecture for Turbo Decoding

The process of decoding data in communication systems is among the most computationally expensive processes. As we move towards cloud based architectures, there is a need for faster data decoding. We propose a novel Sparse Dense Neural Network (SDNN) solution for the turbo decoder module. We show that our proposed solution outperforms both the conventional decoders and the other ML-based decoding algorithms in the literature, in terms of computational complexity. Specifically, a computational gain of about 10% to 20% is achieved, with the BER performance kept comparable to ML-based state-of-the-art turbo decoders. We then show that our solution is also scalable with the order of modulation scheme.