Data-Driven Resource Allocation for Deep Learning in IoT Networks

We consider an Internet of Things (IoT) network, where a large amount of sensor data is transmitted from wireless IoT devices to a central server for the classification of system behaviors. When the number of IoT devices and their sensors is very large, the redundancy in the collected data at the server can also be very high due to the high correlation between the data. By compressing the sensor data at IoT devices for reducing the redundancy, the transmit power can be largely reduced but the accuracy of the classification can also be decreased. Our objective is to maximize the classification accuracy by determining the compression ratio at each device, under the total transmit power constraint. In traditional resource allocation for multiuser wireless networks, the classification accuracy cannot be directly considered since it is usually data-agnostic. On the other hand, in deep learning, the classification accuracy can be maximized by training a neural network architecture using datasets. The deep learning, however, does not usually consider the wireless channel information, that is essential for the resource allocation. In this article, we propose a resource allocation scheme based on a binarized neural network in order to maximize the classification accuracy at the server while satisfying the total transmit power constraint by exploiting both the wireless channel state information and the data-driven approach. In experimental results, we show that the classification accuracy can largely be increased by the proposed scheme using MNIST and CIFAR-10 datasets.