Utilizing Multiple-Access Communication to Realize FNN Computation in Wireless Networks

Driven by the visions of the Internet of Things, more and more devices will join the network and generate large amounts of data. To analyze these data and provide useful information for intelligent applications, artificial neural networks are expected to be widely configured in data centers. However, it incurs excessive latency for the centers to aggregate data from massive distributed devices. To tackle this challenge, we propose wireless-network-based feedforward neural network (WN-FNN), a novel design that utilizes multiple-access communication to perform FNN computation in wireless networks. Based on a technology known as computation over multiple-access channel, the proposed design integrates communication and computation, which dramatically reduces the data aggregation latency. Furthermore, we analyze the impacts of the input noise and received noise on WN-FNN. A regularization strategy is proposed to improve the denoising ability of WN-FNN. An indoor localization application is simulated to verify the performances of WN-FNN and the regularization strategy.