Lifelong Adaptive Machine Learning for Sensor-Based Human Activity Recognition Using Prototypical Networks.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Human Activity Recognition

Background:

• Continual learning (CL) is crucial for dynamic adaptation in human activity recognition (HAR).
• Existing CL methods for HAR often rely on known task boundaries, limiting real-world applicability.
• Current research inadequately addresses task-free, data-incremental learning in HAR.

Purpose of the Study:

• To develop a lifelong adaptive learning framework for sensor-based HAR data streams.
• To address the challenge of catastrophic forgetting in continual learning settings.
• To enable dynamic adaptation to evolving human behaviors without predefined task information.

Main Methods:

• Proposed LAPNet-HAR framework utilizing Prototypical Networks.
• Employed experience replay and continual prototype adaptation to mitigate forgetting.
• Utilized contrastive loss for enhanced inter-class separation in online learning.

Main Results:

• Demonstrated the effectiveness of LAPNet-HAR in task-free, data-incremental continual learning.
• Evaluated performance on five public activity datasets, showing ability to acquire new knowledge while retaining old.
• Empirical results validate the framework's capability in dynamic HAR scenarios.

Conclusions:

• LAPNet-HAR offers a robust solution for continual learning in human activity recognition.
• The task-free, data-incremental approach enhances the practical applicability of CL in real-world HAR systems.
• Findings provide valuable insights for future research in adaptive lifelong learning for HAR.