Training Neural Networks in Cloud Environments: A Methodology and a Comparative Analysis

Abstract

Deep neural networks are solutions to problems that involve pattern recognition, and various works seek to optimize the performance of these networks. This optimization requires suitable hardware, which can be expensive for small and medium organizations. This work proposes a methodology to evaluate the performance and cost of training deep neural networks by assessing how much impact factors such as environment setup, frameworks, and datasets can have on training time and, along with this task, evaluating the total financial cost of the environment for the training process. Experiments were performed to measure and compare the performance and cost of training deep neural networks on cloud platforms such as Azure, AWS, and Google Cloud. In this sense, factors such as the size of the input image and the network architecture significantly impact the training time metric and the total cost.