On the Performance of Deep Learning in the Full Edge and the Full Cloud Architectures

Abstract

Deep learning today surpasses various machine learning approaches in performance and is widely used for variety of different tasks. Deep learning has increased accuracy compared to other approaches for tasks like language translation and image recognition. However, training a deep learning model on a large dataset is a challenging and expensive task that can be time consuming and require large computational resources. Therefore, Different architectures have been proposed for the implementation of deep learning models in machine vision systems to deal with this problem. Currently, the application of deep learning in the cloud is the most common and typical method. Nevertheless, the challenge of having to move the data from where it is generated to a cloud data center so that it can be used to prepare and develop machine learning models represents a major limitation of this approach. As a result, it is becoming increasingly important to consider moving aspects of deep learning to the edge, instead of the cloud, especially with the rapid increase in data volumes and the growing need to act in real time. From this perspective, a comparative study between the full edge and the full cloud architectures based on the performance of the deep learning models implemented in both architectures is elaborated. The results of this study lead us to specify the strengths of both the cloud and the edge for deploying deep learning models, and to choose the optimal architecture to deal with the rapid increase in data volumes and the growing need for real-time action.