Pruning Growing Self-Organizing Map Network for Human Physical Activity Identification.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Wearable Technology

Background:

• Human physical activity identification using wearable sensors is crucial for health analysis.
• Current machine learning models often require extensive labeled data, limiting their generalization ability.
• Difficulty in acquiring large labeled datasets hinders the development of robust activity recognition systems.

Purpose of the Study:

• To propose an unsupervised model, the Pruning Growing Self-Organizing Map (SOM), to overcome limitations of small labeled datasets in human physical activity identification.
• To reduce dependency on labeled data by utilizing an unsupervised training stage followed by minimal labeling of neurons.
• To enhance model generalization and identification accuracy, particularly for underrepresented action categories.

Main Methods:

• Developed a Pruning Growing SOM model for human physical activity identification.
• Employed an unsupervised training phase to allow the model to learn data distribution.
• Incorporated a pruning mechanism to remove inactive neurons, optimizing model scale and consistency with data.

Main Results:

• The Pruning Growing SOM model demonstrates improved identification accuracy, especially on unbalanced datasets.
• The pruning mechanism enhances model consistency with data distribution, leading to better recognition of difficult action categories.
• Model scale is controlled through pruning, resulting in faster inference speeds.

Conclusions:

• The proposed Pruning Growing SOM effectively addresses the challenge of limited labeled data in human physical activity recognition.
• The unsupervised approach combined with pruning offers a more robust and efficient solution for health analysis applications.
• This method significantly improves model performance and inference speed, making it suitable for real-world wearable sensor data.