Optimizing Accuracy-Efficiency Trade-Offs of On-Device Activity Inference With Star Operation

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Lightweight Convolutional Neural Networks (CNNs) excel in sensor-based human activity recognition (HAR) on edge devices due to fast inference.
• Limited convolutional kernel range restricts capturing long-range dependencies in time series sensor data, hindering performance.
• Self-attention mechanisms improve long-range dependency handling but increase computational cost, impacting real-time HAR.

Purpose of the Study:

• To introduce a novel learning paradigm and a computationally efficient Dual-Branch High-Order Interactions (DbHoi) block for mobile HAR.
• To enable lightweight CNNs to capture high-dimensional, non-linear features from sensor data without significant computational overhead.
• To improve the performance of HAR on resource-constrained edge devices.

Main Methods:

• Developed a novel 'star operation' learning paradigm.
• Proposed the lightweight Dual-Branch High-Order Interactions (DbHoi) block.
• Operated in a low-dimensional feature space, analogous to polynomial kernel tricks, to manage computational complexity.
• Conducted experiments on UCI-HAR, UniMiB-SHAR, and OPPORTUNITY benchmarks.

Main Results:

• The proposed DbHoi block consistently outperformed established lightweight networks like MobileNet, ShuffleNet, and GhostNet.
• DbHoi effectively captures complex, high-dimensional features from raw sensor inputs with low computational cost.
• On-device latency analyses confirmed the practical efficiency of DbHoi for real-world HAR applications.

Conclusions:

• The DbHoi block offers a computationally friendly and effective solution for sensor-based HAR on edge devices.
• The 'star operation' paradigm provides a method to implicitly learn complex features without incurring additional computational burden.
• DbHoi demonstrates significant practical merit for real-world human activity recognition deployments.