Towards Learning Discrete Representations via Self-Supervision for Wearables-Based Human Activity Recognition.

Area of Science:

• Computer Science
• Machine Learning
• Wearable Computing

Background:

• Human activity recognition (HAR) traditionally uses continuous sensor data features.
• Past discretization methods for HAR suffered from significant precision loss.
• Advancements in vector quantization (VQ) offer new possibilities for discrete representations.

Purpose of the Study:

• To re-evaluate and apply discretization techniques for HAR using modern methods.
• To demonstrate the effectiveness of learned discrete representations derived from sensor data.
• To explore applications of discrete HAR beyond simple activity classification.

Main Methods:

• Applied recent advancements in vector quantization (VQ) to wearable sensor data.
• Developed a method to learn a direct mapping from sensor data spans to a codebook index.
• Evaluated the approach on a suite of benchmark wearable-based HAR tasks.

Main Results:

• Achieved human activity recognition performance on par with, and often surpassing, continuous methods.
• Demonstrated the potential of learned discretization for wearable applications.
• Showcased the viability of discrete representations for advanced symbolic sequence analysis.

Conclusions:

• Learned discretization via VQ is a viable and effective approach for HAR.
• Discrete representations unlock new analytical tools, similar to those in natural language processing.
• This work signifies a potential paradigm shift in analyzing sensor data for HAR.