Wireless neural network: Enabling neural computing over wireless sensor network based on superposition transmissions

Wireless sensor network (WSN) is a key enabling technology for Internet of Things (IoT), where the sensed data reported by the distributed sensors are transmitted to a core node for intelligent computation and decision. However, the isolation between wireless communication and computing leads to a waste of radio resources, since not all sensed data are required for making a precise enough decision. Hence, we propose a wireless neural network (WNN) to integrate the neural computing and wireless communication by exploiting the superposition characteristics of radio channels as well as the reciprocity between deep artificial neural network and multi-tier WSN. The learning ability of WNN is further enhanced by introducing multi-carrier transmission where the transmit gain of each sub-carrier can be freely trained to increase the number of adjustable network parameters. Experiments on some datasets demonstrate that, similar decision accuracy can be achieved compared with the conventional isolated method, while the radio resource consumption can be greatly reduced due to superposition transmissions.