Explaining and Visualizing Embeddings of One-Dimensional Convolutional Models in Human Activity Recognition Tasks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Human Activity Recognition (HAR) is crucial for various applications.
• One-Dimensional Convolutional Neural Networks (1D CNNs) are effective for HAR but lack interpretability.
• Explaining the features learned by 1D CNNs is a significant challenge.

Purpose of the Study:

• To develop a novel eXplainable Artificial Intelligence (XAI) method for visualizing 1D CNN features in HAR.
• To provide insights into the decision-making process of 1D CNNs using t-Distributed Stochastic Neighbor Embedding (t-SNE).
• To enhance the reliability and practicality of 1D CNN models in real-world HAR scenarios.

Main Methods:

• Utilized t-Distributed Stochastic Neighbor Embedding (t-SNE) for feature visualization.
• Applied the XAI method to the deepest layer of 1D CNNs before classification.
• Trained and evaluated 1D CNN models on public HAR datasets (SHO and HAPT).

Main Results:

• Demonstrated that features learned from one dataset can generalize to others.
• Achieved high performance: 0.98 accuracy on the SHO dataset and 0.93 accuracy on the HAPT dataset.
• Visualizations provided insights into model decision-making and potential bias.

Conclusions:

• The proposed XAI method effectively visualizes 1D CNN features for HAR.
• The approach enhances model interpretability and aids in detecting bias or explaining errors.
• This work advances the application of XAI in deep learning for HAR.