On-Device Deep Learning Inference for Efficient Activity Data Collection.

Area of Science:

• Computer Science
• Machine Learning
• Human-Computer Interaction

Background:

• Accurate labeling of activity data is crucial for human activity recognition (HAR) systems.
• User motivation is key for collecting sufficient and high-quality activity annotations.
• On-device deep learning offers potential for context-aware applications on mobile devices.

Purpose of the Study:

• To develop and evaluate a novel method for alleviating labeling effort in HAR systems.
• To leverage on-device deep learning inference for motivating users in activity data collection.
• To improve both data quality and quantity for enhanced HAR system performance.

Main Methods:

• Utilized a long short-term memory (LSTM)-based deep learning model for on-device activity inference.
• Implemented a feedback mechanism where estimated activities motivate users to provide accurate labels.
• Compared the proposed method against a traditional notification-based labeling approach using smartphone sensors.

Main Results:

• The proposed method demonstrated significant improvements in data quality, evidenced by better classification model performance.
• The study also showed an increase in data quantity, indicated by a higher number of collected data points.
• On-device inference effectively enhanced user motivation and engagement in the labeling process.

Conclusions:

• On-device deep learning inference, particularly with LSTM, can effectively reduce labeling effort and improve data collection for HAR systems.
• The proposed feedback mechanism shows promise for motivating users and increasing data quantity and quality.
• This work contributes to advancing HAR systems through efficient and user-centric data collection strategies and releases a preliminary dataset.