Neural Architecture Search for Resource-Constrained Internet of Things Devices

Abstract

The traditional process of extracting knowledge from the Internet of Things (IoT) happens through Cloud Computing by offloading the data generated in the IoT device to processing in the cloud. However, this regime significantly increases data transmission and monetary costs and may have privacy issues. Therefore, it is paramount to find solutions that achieve good results and can be processed as close as possible to an IoT object. In this scenario, we developed a Neural Architecture Search (NAS) solution to generate models small enough to be deployed to IoT devices without significantly losing inference performance. We based our approach on Evolutionary Algorithms, such as Grammatical Evolution and NSGA-II. Using model size and accuracy as fitness, our proposal generated a Convolutional Neural Network model with less than 2 MB, achieving an accuracy of about 81 % in the CIFAR-10 and 99 % in MNIST, with only 150 thousand parameters approximately.