FreqSense: Adaptive Sampling Rates for Sensor-Based Human Activity Recognition Under Tunable Computational Budgets.

Area of Science:

• Computer Science
• Machine Learning
• Signal Processing

Background:

• Deep convolutional networks achieve high accuracy in sensor-based human activity recognition (HAR).
• Practical HAR deployment is hindered by variable computational demands for reliable predictions.
• Existing methods often require fixed computational budgets, regardless of sample complexity.

Purpose of the Study:

• To develop an adaptive inference method for HAR that optimizes computational resource allocation.
• To leverage signal frequency characteristics for efficient activity recognition.
• To reduce computational cost without sacrificing prediction accuracy.

Main Methods:

• Proposed an adaptive resolution network combining subsampling and conditional early-exit strategies.
• Implemented a multi-resolution subnetwork architecture where easier samples are classified by lower-resolution subnetworks.
• Dynamically selected sampling rates based on confidence thresholds, enabling early termination for simple activities.

Main Results:

• Demonstrated the effectiveness of the adaptive resolution network across four diverse HAR benchmark datasets.
• Achieved a favorable accuracy-cost tradeoff, adapting computational load to sample complexity.
• Benchmarked average latency on real hardware, validating practical performance.

Conclusions:

• The proposed adaptive resolution network offers an efficient approach to sensor-based HAR.
• Dynamically adjusting computational effort based on signal frequency and sample difficulty optimizes performance.
• This method provides a flexible solution for HAR systems with varying computational budgets.