HARNAS: Human Activity Recognition Based on Automatic Neural Architecture Search Using Evolutionary Algorithms.

Area of Science:

• Computer Science
• Machine Learning
• Wearable Technology

Background:

• Human activity recognition (HAR) using wearable sensors is crucial but limited by device resources.
• Lightweight deep learning architectures are needed for efficient HAR.
• Neural Architecture Search (NAS) offers automated model design to minimize human intervention.

Purpose of the Study:

• To propose HARNAS, an NAS approach for designing efficient HAR models.
• To optimize models for both performance (F1 score) and computational efficiency (FLOPs, MAC).
• To evaluate HARNAS's effectiveness and portability on benchmark datasets.

Main Methods:

• Utilized NSGA-II as the multi-objective search strategy within HARNAS.
• Formulated HAR as a tri-objective optimization problem considering F1 score, FLOPs, and MAC.
• Trained and evaluated models on the Opportunity and UniMiB-SHAR datasets.

Main Results:

• HARNAS automatically designed models that surpassed manually optimized counterparts.
• Achieved an F1 score of 92.16% with a model size of 0.32 MB on the Opportunity dataset.
• Demonstrated the portability and effectiveness of the NAS-generated models.

Conclusions:

• HARNAS successfully automates the design of high-performance, lightweight HAR models.
• The approach offers a significant advancement over manual model tuning for wearable applications.
• Automated NAS provides a viable solution for resource-constrained HAR systems.