Sensor-Modality-Aware Human Activity Recognition with the Convolutional Tsetlin Machine: Interpretable and...

Area of Science:

• Machine Learning
• Sensor Data Analysis
• Artificial Intelligence

Background:

• Human activity recognition (HAR) traditionally uses complex models with limited interpretability and high resource demands.
• Existing methods struggle with transparency, verifiability, and deployment on resource-constrained devices.

Purpose of the Study:

• Investigate the Convolutional Tsetlin Machine (CTM) for multimodal HAR using raw inertial signals.
• Evaluate CTM's performance, interpretability, and computational efficiency compared to conventional methods.
• Analyze the contribution of different sensor modalities (accelerometer, gyroscope) to HAR accuracy.

Main Methods:

• Utilized the UCI-HAR dataset with raw inertial signals (9x128) instead of pre-computed features.
• Trained separate CTM classifiers for various combinations of sensor modalities (accelerometer, gyroscope) and signal groups.
• Decomposed inertial signals by sensor source, signal group, and coordinate axis for systematic analysis.

Main Results:

• CTM achieved strong predictive performance for HAR using raw inertial data.
• The neuro-symbolic nature of CTM provided interpretable, logic-based decision rules.
• CTM demonstrated low computational complexity, suitable for efficient, low-power on-device learning.

Conclusions:

• CTM offers a principled framework for explainable, resource-efficient, and deployable HAR systems.
• The study highlights CTM's potential for trustworthy multimodal sensing by balancing performance, interpretability, and efficiency.
• CTM's ability to learn from raw data and provide transparent rules enhances its suitability for practical embedded applications.