Robust Feature Representation Using Multi-Task Learning for Human Activity Recognition.

Area of Science:

• Computer Science
• Machine Learning
• Signal Processing

Background:

• Human Activity Recognition (HAR) relies on learning patterns from sensory data for effective generalization.
• Representation learning is key to address similar activity patterns and inter-subject variations in HAR.
• Existing methods struggle with generalization to unseen data and users.

Purpose of the Study:

• To develop a robust representation learning method for HAR that improves generalization to unseen data and users.
• To enhance the model's ability to learn underlying factors from sensor signals.
• To investigate the impact of activation functions on signal reconstruction in HAR.

Main Methods:

• A novel multi-channel asymmetric auto-encoder was developed for precise signal reconstruction and unsupervised feature extraction.
• The study proposed a multi-task learning framework integrating signal reconstruction and HAR tasks.
• Shared layers were utilized to learn common features between the two tasks, enhancing representation learning.

Main Results:

• The multi-task learning model achieved high accuracy across multiple public HAR datasets (UCI-HAR, MHealth, PAMAP2, USC-HAD) and an in-house dataset.
• Accuracies ranged from 88% to 99% on tested datasets, demonstrating consistent performance.
• The model exhibited strong generalization capabilities, particularly for users not included in the training phase.

Conclusions:

• The proposed multi-task learning approach effectively enhances representation learning for HAR.
• The method improves model robustness and generalization by learning shared underlying factors from sensor data.
• This technique offers a promising solution for more reliable and accurate Human Activity Recognition systems.