Transfer learning for wearable computers

Wearable sensors are taking a bold stance in becoming the principal component of monitoring systems with wide applications in health monitoring, assisted living, sport studies, entertainment, and diet monitoring. Various wearable sensors with different sensing capabilities including smartwatches, smartphones, wrist-band sensors, sports shoes, and sensors embedded in clothing have been used for the aforementioned applications. However, in the presence of the user’s diverse preference and requirement of various environments, changes in the configuration and type of sensors are highly possible. For example, a user who has been using a smartphone for a while may acquire a new smartwatch. Besides changes in the sensor modalities and configurations, the changes in the user characteristics are highly possible. A system trained on the data of specific users should be able to work for a new user with different characteristics such as different body mass index, skin tone, physiological attributes, and even sensor placement preferences. As a result of such changes in the environment, the machine learning and signal processing components should be updated or retrained. Otherwise, the performance of the models designed on the old training data will be significantly degraded when those models are used with the data collected under a new configuration. The main challenge with retraining the underlying signal processing and machine learning models is that it requires collecting a sufficiently large amount of labeled training data. The process of collecting such data is very challenging and overwhelming if the users are asked to provide them. Therefore, it is of paramount interest to transfer the knowledge from an old domain to a new domain, with minimum effort required by the users. In dynamic environments, where the configuration of the wearable systems can constantly change over time, the notion of transfer learning for these devices becomes tremendously important.