Experimentation and Analysis of Ensemble Deep Learning in IoT Applications

This paper presents an experimental study of Ensemble Deep Learning (DL) techniques for the analysis of time series data on IoT devices. We have shown in our earlier work that DL demonstrates superior performance compared to traditional machine learning techniques on fall detection applications due to the fact that important features in time series data can be learned and need not be determined manually by the domain expert. However, DL networks generally require large datasets for training. In the health care domain, such as the real-time smartwatch-based fall detection, there are no publicly available large annotated datasets that can be used for training, due to the nature of the problem (i.e. a fall is not a common event). Moreover, fall data is also inherently noisy since motions generated by the wrist-worn smartwatch can be mistaken for a fall. This paper explores combing DL (Recurrent Neural Network) with ensemble techniques (Stacking and AdaBoosting) using a fall detection application as a case study. We conducted a series of experiments using two different datasets of simulated falls for training various ensemble models. Our results show that an ensemble of deep learning models combined by the stacking ensemble technique, outperforms a single deep learning model trained on the same data samples, and thus, may be better suited for small-size datasets.