ESAI: Efficient Split Artificial Intelligence via Early Exiting Using Neural Architecture Search

Abstract

Deep neural networks have demonstrated superior performance in various computer vision tasks compared to traditional machine learning algorithms. However, deploying these models on resource-constrained mobile and IoT devices poses computational challenges. Many devices resort to cloud computing, where complex deep learning models analyze data on servers. This approach increases communication costs and hampers system efficiency in the absence of a network connection. In this paper, we introduce a novel framework for deploying deep neural networks on IoT devices. This framework leverages both cloud and on-device models by extracting meta-information from each sample's classification result. It assesses the classification's performance to determine whether sending the sample to the server is necessary. Extensive experiments on CIFAR10 and CINIC10 datasets reveal that only 45% of CIFAR10 and 60% of CINIC10 test data need to be transmitted to the server using this technique. The overall accuracy of the framework is 94% and 89%, respectively, enhancing the accuracy of both client and server models.