Few-Shot Human Activity Recognition on Noisy Wearable Sensor Data

Abstract

Most existing wearable sensor-based human activity recognition (HAR) models are trained on substantial labeled data. It is difficult for HAR to learn new-class activities unseen during training from a few samples. Very few researches of few-shot learning (FSL) have been done in HAR to address the above problem, though FSL has been widely used in computer vision tasks. Besides, it is impractical to annotate sensor data with accurate activity labels in real-life applications. The noisy labels have great negative effects on FSL due to the limited samples. The weakly supervised few-shot learning in HAR is challenging, significant but rarely researched in existing literature. In this paper, we propose an end-to-end Weakly supervised Prototypical Networks (WPN) to learn more latent information from noisy data with multiple instance learning (MIL). In MIL, the noisy instances (subsequences of segmentation) have different labels from the bag’s (segmentation’s) label. The prototype is the center of the instances in WPN rather than less discriminative bags, which determines the bag-level classification accuracy. To get the most representative instance-level prototype, we propose two strategies to refine the prototype by selecting high-probability instances same as their bag’s label iteratively based on the distance-metric. The model is trained by minimizing the instance-level loss function and infers the final bag-level labels from instance-level labels. In the experiments, our proposals outperform existing approaches and achieve higher average ranks.