Combining Neural Architecture Search and Weight Reshaping for Optimized Embedded Classifiers in Multisensory Glove.

Area of Science:

• Artificial Intelligence
• Embedded Systems
• Wearable Technology

Background:

• Intelligent sensing systems are crucial for wearable devices in robotics and human-machine interaction.
• Energy autonomy is a key challenge for these systems due to resource constraints.
• Optimizing embedded neural networks for accuracy and efficiency is essential.

Purpose of the Study:

• To develop energy-efficient classifiers for multisensory gloves by integrating Hardware-Aware Neural Architecture Search (HW-NAS) with optimization techniques.
• To balance accuracy and efficiency in embedded neural networks for resource-constrained wearable devices.
• To demonstrate the effectiveness of combined HW-NAS and optimization for energy-autonomous systems.

Main Methods:

• Hardware-Aware Neural Architecture Search (HW-NAS) was employed to automatically derive 1D-CNN models suitable for the NUCLEO-F401RE board.
• Model optimization techniques, including weight reshaping and quantization, were applied to further reduce resource usage.
• The integrated approach was evaluated on three datasets using a multisensory glove.

Main Results:

• The optimized models achieved improved classification accuracy compared to NAS-only baselines.
• Significant reductions were observed in inference time (average 75%), flash memory (average 69%), and RAM (average 45%).
• The combined approach effectively reduced model size, memory footprint, and latency.

Conclusions:

• Integrating HW-NAS with optimization techniques is highly effective for developing efficient classifiers for wearable devices.
• This approach paves the way for creating practical, energy-autonomous wearable sensing systems.
• The study demonstrates a viable strategy for optimizing embedded neural networks under strict resource limitations.