Aprendizado offline e online de redes neurais no contexto de casas inteligentes e de computação em névoa

Abstract

As smart applications based on Artificial Neural Networks (ANNs) become highly popular, particularly the models comprising deep learning, some drawbacks of traditional cloud-based deployment emerge. Issues like high monetary cost, for storing and running applications, low privacy on data and models, and high latency experienced by cloud-based neural networks might make their use difficult, leading to poor user experiences. Fog computing appears therefore as an interesting alternative. This work explores, in the context of smart homes, a fog topology as an alternative to online learning and ANN-based models running offline. The work proposes using different shallow components to form deeper models; it also adopts a traditional deep recursive approach to deal with temporal aspects of data. Experiments involving offline learning compare their performance on eight different classification problems which consist of activities performed by a user in each one out of eight rooms in the smart home addressed as the case study. Results show that the hybridization of an auto-encoder with classifiers based on multi-layer perceptrons can detect rare activities and provide good results for almost all rooms, particularly when encompassing suitable neural structure sizes in the pipelines. However, it is worth mentioning that the traditional multilayer model is quite competitive. In the online context, although the performance of the best approach decreases, as expected, some relevant insights result from experiments, especially that fog computing provides results not too far from cloud systems, yet demanding fewer resources. The proposal based on fog computing and online learning appears therefore as an alternative when dealing with streaming data on restricted environments in terms of computation resources or time.