Parameter-Efficient Deep Learning for Ultrasound-Based Human-Machine Interfaces

A research team from KU Leuven and IMEC has published a systematic study on parameter-efficient deep learning for ultrasound-based human-machine interfaces (HMIs). Their work, submitted to ICPR 2026, addresses a gap in the literature by comparing six different deep learning models on the only publicly available benchmark, the Ultrasound Adaptive Prosthetic Control (Ultra-Pro) dataset. The goal is to enable reliable Hand Pose Estimation (HPE), which could unlock interfaces supporting up to 23 degrees of freedom for intuitive, rich interaction with machines and prosthetics.

Their key finding is that a custom 4-layer model, dubbed UDACNN, outperforms more complex architectures when paired with optimal preprocessing and training. By using the envelope of the raw radio-frequency (RF) ultrasound signals as input and a step learning rate scheduler, UDACNN achieved a benchmark accuracy of 77.72%. This constitutes a 0.88% absolute improvement over previous baselines and, more impressively, surpasses the performance of the XceptionTime model by 2.28 percentage points while featuring 87.52% fewer parameters.

The study underscores that model efficiency is not solely about architecture size but the 'appropriate combination of model, preprocessing and training algorithm.' This parameter-efficient approach is crucial for deploying such systems on edge devices with limited computational resources, like wearable sensors or embedded systems in prosthetics. The work provides a reproducible framework and benchmark for future development in a field poised to create more natural and capable interfaces for control, rehabilitation, and augmented reality.