Generalized Few-Shot Learning For Wearable Sensor-based Human Activity Recognition

Abstract

Few-Shot Learning (FSL) aims to create a classifier which generalizes to classes not present in the training set given just a few samples from each new class. Generalized FSL recognizes samples from classes that are both present and not present in the training set. We developed the first generalized FSL system for Human Activity Recognition (HAR) based on data from wearable sensors. This enabled the classification of new human activities without the high cost of collecting data and allowed for increased personal variation of performing certain activities. We implemented prototypical networks and a center loss based model for FSL. We trained and evaluated two additional classifiers. The first one recognizes source classes and the second determines if a target is from source/target domain. We evaluated our model on three publicly available datasets (UTWENTE, PAMAP2 and OPP) and showed that our methods significantly outperformed the state-of-the-art on the FSL task.