Contrastive Predictive Coding for Human Activity Recognition

Feature extraction is crucial for human activity recognition (HAR) using body-worn movement sensors. Recently, learned representations have been used successfully, offering promising alternatives to manually engineered features. Our work focuses on effective use of small amounts labeled data and the opportunistic exploitation unlabeled that are straightforward collect in mobile ubiquitous computing scenarios. We hypothesize demonstrate explicitly considering temporality sensor at representation level plays an important role HAR challenging introduce Contrastive Predictive Coding (CPC) framework recognition, which captures temporal structure streams. Through a range experimental evaluations real-life tasks, we its effectiveness improved HAR. CPC-based pre-training self-supervised, resulting can be integrated into standard chains. It leads significantly performance when only training available, thereby demonstrating practical value our approach. series experiments, also develop guidelines help practitioners adapt modify towards other