Exploring Regularization Methods for Domain Generalization in Accelerometer-Based Human Activity Recognition.

Area of Science:

• Machine Learning
• Human Activity Recognition
• Computer Vision

Background:

• Domain Generalization (DG) is crucial in Machine Learning (ML).
• Human Activity Recognition (HAR) presents inherent domain diversity, making it ideal for DG research.
• Bridging the generalization gap between traditional and deep learning models is an ongoing challenge.

Purpose of the Study:

• Investigate regularization methods for improving DG in HAR.
• Compare deep learning models with handcrafted features in Out-of-Distribution (OOD) settings.
• Assess the effectiveness of sparse training, Mixup, Distributionally Robust Optimization (DRO), and Sharpness-Aware Minimization (SAM).

Main Methods:

• Applied regularization techniques (sparse training, Mixup, DRO, SAM) to deep learning models.
• Evaluated model performance in OOD settings across multiple domains.
• Utilized homogenized public datasets for consistent comparison.
• Compared deep learning approaches against traditional models with handcrafted features.

Main Results:

• Mixup and SAM demonstrated the strongest performance among the tested regularizers.
• Despite improvements, regularized deep learning models did not surpass handcrafted feature-based models.
• Regularization techniques offered partial improvements in OOD robustness.

Conclusions:

• Regularization methods can enhance OOD robustness in HAR to a degree.
• Handcrafted features remain superior for achieving domain generalization in HAR tasks.
• Further research may be needed to fully leverage deep learning for HAR domain generalization.