Serial and Parallel Convolutional Neural Network Designs for NFDM Signals

Two conceptual convolutional neural network designs are proposed for directly decoding NFDM signals with the consideration of hardware implementation. A serial network is designed for small network size suitable for small user applications and a parallel network is designed for speed suitable for places such as data centres. In the numerical demonstrations, the serial design only occupies 0.5 MB of memory space while the parallel design occupies 128 MB of memory but allow parallel computing. Both network designs were trained with simulated data and able to reach more than 99.9% accuracy.