Learning Using Privileged Information for Wearable-based Human Activity Recognition

Most sensor-based Human Activity Recognition (HAR) approaches assume that the same set of sensors is available for both training and testing. Existing public HAR datasets collected in controlled environments often contain multiple and various types of sensors, and HAR models are often trained using all or most of them. However, expecting people to wear multiple sensors in everyday life may not be realistic, and only a subset of those sensors may only be available in practice. It would be beneficial to use the extra sensors, only available at training time, to train better-performing HAR models. Such a learning paradigm is often referred to as Learning Using Privileged Information (LUPI). In this research project, we will explore different LUPI approaches and identify in which situations LUPI can be beneficial for wearable-based HAR.