Data and resource aware incremental ML training in support of pervasive applications

Abstract

Nowadays, the use of Artificial Intelligence (AI) and Machine Learning (ML) algorithms is increasingly affecting the performance of innovative systems. At the same time, the advent of the Internet of Things (IoT) and the Edge Computing (EC) as means to place computational resources close to users create the need for new models in the training process of ML schemes due to the limited computational capabilities of the devices/nodes placed there. In any case, we should not forget that IoT devices or EC nodes exhibit less capabilities than the Cloud back end that could be adopted for a more complex training upon vast volumes of data. The ideal case is to have, at least, basic training capabilities at the IoT-EC ecosystem in order to reduce the latency and face the needs of near real time applications. In this paper, we are motivated by this need and propose a model that tries to save time in the training process by focusing on the training dataset and its statistical description. We do not dive into the architecture of any ML model as we target to provide a more generic scheme that can be applied upon any ML module. We monitor the statistics of the training dataset and the loss during the process and identify if there is a potential to stop it when not significant contribution is foreseen for the data not yet adopted in the model. We argue that our approach can be applied only when a negligibly decreased accuracy is acceptable by the application gaining time and resources from the training process. We provide two algorithms for applying this approach and an extensive experimental evaluation upon multiple supervised ML models to reveal the benefits of the proposed scheme and its constraints.