Earables have emerged as a unique platform for ubiquitous computing by augmenting ear-worn devices with state-of-the-art sensing. This new platform has spurred a wealth of new research exploring what can be detected on a wearable, small form factor. As a sensing platform, the ears are less susceptible to motion artifacts and are located in close proximity to a number of important anatomical structures including the brain, blood vessels, and facial muscles which reveal a wealth of information. They can be easily reached by the hands and the ear canal itself is affected by mouth, face, and head movements. We have conducted a systematic literature review of 271 earable publications from the ACM and IEEE libraries. These were synthesized into an open-ended taxonomy of 47 different phenomena that can be sensed in, on, or around the ear. Through analysis, we identify 13 fundamental phenomena from which all other phenomena can be derived, and discuss the different sensors and sensing principles used to detect them. We comprehensively review the phenomena in four main areas of (i) physiological monitoring and health, (ii) movement and activity, (iii) interaction, and (iv) authentication and identification. This breadth highlights the potential that earables have to offer as a ubiquitous, general-purpose platform.

Deep learning models have shown excellent performance in human activity recognition tasks. However, these models typically require large amounts of computational resources, which makes them inefficient to deploy on edge devices. Furthermore, the superior performance of deep learning models relies heavily on the availability of large datasets to avoid over-fitting. However, the expensive efforts for labeling limits the amount of datasets. We address both challenges by designing a more lightweight model, called TinyHAR. TinyHAR is designed specifically for human activity recognition employing different saliency of multi modalities, multimodal collaboration, and temporal information extraction. Initial experimental results show that TinyHAR is several times smaller and often meets or even surpasses the performance of DeepConvLSTM, a state-of-the-art human activity recognition model.

Sensor-based HAR models face challenges in cross-subject generalization due to the complexities of data collection and annotation, impacting the size and representativeness of datasets. While data augmentation has been successfully employed in domains like natural language and image processing, its application in HAR remains underexplored. This study presents AutoAugHAR, an innovative two-stage gradient-based data augmentation optimization framework. AutoAugHAR is designed to take into account the unique attributes of candidate augmentation operations and the unique nature and challenges of HAR tasks. Notably, it optimizes the augmentation pipeline during HAR model training without substantially extending the training duration. In evaluations on eight inertial-measurement-units-based benchmark datasets using five HAR models, AutoAugHAR has demonstrated superior robustness and effectiveness compared to other leading data augmentation frameworks. A salient feature of AutoAugHAR is its model-agnostic design, allowing for its seamless integration with any HAR model without the need for structural modifications. Furthermore, we also demonstrate the generalizability and flexible extensibility of AutoAugHAR on four datasets from other adjacent domains. We strongly recommend its integration as a standard protocol in HAR model training and will release it as an open-source tool1.

Neural network models have demonstrated exceptional performance in wearable human activity recognition (HAR) tasks. However, the increasing size or complexity of HAR models significantly impacts their deployment on wearable devices with limited computational power. In this study, we introduce a novel HAR model architecture named Multi-Layer Perceptron-HAR (MLP-HAR), which contains solely fully connected layers. This model is specifically designed to address the unique characteristics of HAR tasks, such as multi-modality interaction and global temporal information. The MLP-HAR model employs fully connected layers that alternately operate along the modality and temporal dimensions, enabling multiple fusions of information across these dimensions. Our proposed model demonstrates comparable performance with other state-of-the-art HAR models on six open-source datasets, while utilizing significantly fewer learnable parameters and exhibiting lower model complexity. Specifically, the complexity of our model is at least ten times smaller than that of the TinyHAR model and several hundred times smaller than the benchmark model DeepConvLSTM. Additionally, due to its purely fully connected layer-based architecture, MLP-HAR offers the advantage of ease of deployment. To substantiate these claims, we report the inference time performance of MLP-HAR on the Samsung Galaxy Watch 5 PRO and the Arduino Portenta H7 LITE, comparing it against other state-of-the-art HAR models.

Recent trends in Wearable Human Activity Recognition (WHAR) have led to an unprecedented 42.9% increase in scholarly articles in 2022, underscoring the urgency for a comprehensive review to systematically categorize their varied research directions. Moreover, our analysis reveals that the contributions of current articles often deviate from the traditional stages of the human activity recognition pipeline, as established in prior literature. This misalignment suggests the necessity for an updated pipeline that more accurately reflects the intricacies and nuances of WHAR studies. In response, we review WHAR articles from 2021 to 2023 and introduce an innovative WHAR pipeline, emphasizing a research-focused approach. This new pipeline offers distinct advantages: it provides researchers with a clear and systematic categorization of WHAR articles, thereby enhancing understanding of the field. For practitioners, it facilitates the selection of customized methods for each stage, thereby optimizing final assembled model efficacy.

The rapid development of emerging domains, such as the Internet of Things and wearable technologies, necessitates the development of flexible, stretchable, and non-toxic devices that can be manufactured at an ultra-low cost. Printed electronics has emerged as a viable solution by offering not only the aforementioned features but also a high degree of customization, which enables the personalization of products and facilitates the low-cost product development process even in small batches. In the context of printed electronics, printed neuromorphic circuits offer highly customized and bespoke realization of artificial neural networks to achieve desired functionality with very small number of hardware components. However, since analog components are utilized, the performance of printed neuromorphic circuits can be influenced by various factors. In this work, we focus on three main factors that perturb the circuit output from the designed values, namely, variations due to printing errors, aging effects of printed resistors, and input variations originating from sensing uncertainty. In the described approach, these variations are taken into account during the design (training) to ensure the dependability of the printed neuromorphic circuits. With this approach, the expected accuracy and the robustness of printed neural networks can be increased by 27% and 74%, respectively. Moreover, the ablation study suggests that, aging effect and printing variation may have similar effects on the functionality of printed neural networks. In contrast, the impact of sensing uncertainty on printed neural networks is almost orthogonal to aging and printing variations.