Reducing Label Dependency in Human Activity Recognition with Wearables: From Supervised Learning to Novel Weakly Self-Supervised Approaches

Introduction

The recognition of human activities with wearables has advanced significantly in recent years, but it is still faced with several challenges. Among these, the dependency on labels remains a major obstacle to achieving high accuracy.

In this context, a new study presents innovative solutions to reduce label dependency and improve the performance of activity recognition.

Methodology

The researchers developed a weakly supervised framework that combines multiple machine learning techniques, including limited-label weak supervision and multi-tasking. The framework was tested on various activity recognition datasets and demonstrated competitive performance compared to traditional methods.