Robust Machine Learning for Low-Power Wearable Devices: Challenges and Opportunities

Wearable devices are becoming increasingly popular for a number of applications, including health monitoring, rehabilitation, and fitness. More broadly, low-power Internet of Things (IoT) devices are being used for environmental monitoring, wide area sensing, and smart cities. The low-power devices have been enabled through advances in sensor technology, processing, and communication technology. Similarly, potentially transformative applications using wearable devices are enabled through the use of edge machine learning (ML) algorithms. The ML algorithms process the sensor data at runtime to identify and monitor the parameters of interest to the users. To be effective and provide a high quality of service to the users, the ML algorithms must be robust to changing environmental conditions and potential user errors. For instance, the user may change the orientation or position of the sensors due to an error or change in preferences. The ML models must adapt to these changes and provide accurate monitoring. To this end, this chapter reviews the current state of the art and challenges in robust machine learning algorithms for wearable devices. We also perform a case study with a human activity recognition application to analyze the feasibility of using generative networks for robust activity recognition. Overall, this chapter aims to serve as a resource for researchers working at the intersection of wearable devices and edge ML.