Search

The study of Domain Generalization (DG) has gained considerable momentum in the Machine Learning (ML) field. Human Activity Recognition (HAR) inherently encompasses diverse domains (e.g., users, devices, or datasets), rendering it an ideal testbed for exploring Domain Generalization. Building upon recent work, this paper investigates the application of regularization methods to bridge the generalization gap between traditional models based on handcrafted features and deep neural networks. We apply various regularizers, including sparse training, Mixup, Distributionally Robust Optimization (DRO), and Sharpness-Aware Minimization (SAM), to deep learning models and assess their performance in Out-of-Distribution (OOD) settings across multiple domains using homogenized public datasets. Our results show that Mixup and SAM are the best-performing regularizers. However, they are unable to match the performance of models based on handcrafted features. This suggests that while regularization techniques can improve OOD robustness to some extent, handcrafted features remain superior for domain generalization in HAR tasks. [ABSTRACT FROM AUTHOR]

• Thorough analysis of wearable Fall Detection Systems (FDS) based on neural networks. • Existing research has only focused on optimizing the recognition metrics. • The study reveals significant methodological flaws in the evaluation of FDSs. • Important implementability issues are normally neglected by the literature. • Operational and social aspects are also ignored by the related literature. As a result of the challenge posed by demographic aging and the important personal and medical costs associated to falls of elderly people, the study of automatic Fall Detection Systems (FDS) has gained significant research attention during the last decade. Due to the intrinsic difficulty of identifying falls with a deterministic or a simple threshold-based method, Artificial Neural Networks (ANNs) and other machine learning strategies have been largely employed by the literature as the detection algorithms. This paper thoroughly revises and analyzes those papers in the literature that have proposed a FDS that is partly or entirely based on the use of an ANN. The study, which is focused on wearable devices (as they normally impose higher battery and computing restrictions), discusses the most recent trends in the application of neural architectures to the problem of automatic detection. In this regard, the analysis reveals that the related literature has normally emphasized the maximization of the detection performance, while disregarding some methodological, operative implementation-related issues of utmost importance for the assessment of the actual feasibility of these applications. In particular, the investigation shows that cross validating the detection methods with public benchmarking datasets is gaining importance in recent works. Conversely, the practical problems related to the use of wearables (consumption of power and computing resources, configuration and limitations of the sensing units and the wireless transmissions, etc.) are still normally disregarded. [ABSTRACT FROM AUTHOR]