Neural Network OFDM Receiver Design and FPGA Implementation
Neural networks are powerful processors and when trained with large amounts of diverse data appropriately using multi-objective techniques employed in modern machine learning, they exhibit superb generalization performance capabilities in many applications. These models are also extremely suitable for real-time efficient implementations due to their highly regular layered functional structures. We will present and discuss a process of designing and implementing a neural network based OFDM receiver for indoor/outdoor wireless communications, using a combination of multilayer convolutional, dense, and recurrent neural networks to handle channel estimation and equalization, demodulation, and decoding components of the receiver. In particular, the process will follow a curriculum learning approach, breaking the problem into components, training parts of the neural network on each subtask, and combining them for final joint optimization on real data. Real-time FPGA implementation of the final trained neural network will be incorporated into software-defined radios. Training and test results will be based on synthetic data and real data from Arena (indoor) and Coliseum (outdoor) wireless communication experimental setups.