Explaining One-Dimensional Convolutional Models in Human Activity Recognition and Biometric Identification Tasks.

Area of Science:

• Wearable technology and human-computer interaction.
• Machine learning and artificial intelligence.
• Biometrics and pattern recognition.

Background:

• Human activity recognition (HAR) is crucial due to wearable device popularity.
• Deep learning (DL) models are widely used for HAR, with subject-dependent (SD) and subject-independent (SI) validation methods.
• Understanding model decision-making in HAR and biometric user identification (BUI) is essential for reliable performance.

Purpose of the Study:

• To develop visual explanations for DL models in HAR and BUI tasks using accelerometer data.
• To investigate the correlation between HAR and BUI models' decision-making processes.
• To identify potential database biases and improve model generalization capabilities.

Main Methods:

• Adaptation of gradient-weighted class activation mapping (grad-CAM) for 1D convolutional neural networks (CNNs).
• Application of grad-CAM to visualize decision-making in HAR and BUI models.
• Quantitative evaluation of model accuracy for both SD and SI approaches.

Main Results:

• Achieved high accuracy for HAR (0.978 SD, 0.755 SI) and BUI (0.937 average).
• Demonstrated that SD HAR performance benefits from learning individual user signatures, akin to BUI.
• Identified differences in CNN focus: larger segments for BUI, smaller segments for HAR.
• Utilized grad-CAM to detect database biases like signal discontinuities.

Conclusions:

• Combining explainable AI techniques with DL enhances model design and reliability in HAR and BUI.
• Visual explanations help mitigate result overestimation and identify biases.
• Improved model generalization is achievable through explainable methods.