Design and Prototyping of Neural Network Compression for Non-Orthogonal IoT Signals

The non-orthogonal IoT signal, following the bandwidth compression spectrally efficient frequency division multiplexing (SEFDM) characteristics, can bring benefits in enhanced massive device connections, signal coverage extension and data rate increase, but at the cost of computational complexity. Resource-constrained IoT devices have limited memory storage and complex signal processing is not allowed. Machine learning can simplify signal detection by training a general data-driven signal detection model. However, fully connected neural networks would introduce processing latency and extra power consumption. Therefore, the motivation of this work is to investigate different neural network compression schemes for system simplification. Three compression strategies are studied including topology compression, weight compression and quantization compression. These methods show efficient neural network compression with trade-offs between computational complexity and bit error rate (BER) performance. Practical neural network prototyping is evaluated as well on a software defined radio (SDR) platform. Results show that the practical weight compression neural network can achieve similar performance as the fully connected neural network but with great resource saving.