Evaluation of Regularization-based Continual Learning Algorithms in the Context of Human Activity Recognition

Abstract

In recent years, sensor-based human activity recognition (HAR) has gained a lot of attention in continual learning studies. With a growing number of HAR applications deployed in real-world contexts, continual learning plays a vital role in allowing activity models to adapt to the constant changes in activity routines conducted by users. Most of the applications in HAR intend to solve the problems of learning new activity classes or variations in the data distribution of the existing activities within the limited computational resources. Previous studies have applied different regularization-based methods for their efficiency and simplicity in the applications. However, they focused solely on the single method implementation in the HAR domain without studying the strengths and weaknesses of these methods, which leads to the lack of dynamicity in the model training process. In this paper, we provide a comprehensive comparison of three regularization-based methods in the HAR domain exploring their specialization and limitations. Conducted on the UCI HAR dataset, the experiment provides the results that each technique has its own strength and weakness and no single technique outperformed all others in all scenarios considered.