Unsupervised Domain Adaptation for Inter-Session Re-Calibration of Ultrasound-Based HMIs.

Area of Science:

• Human-Computer Interaction
• Machine Learning
• Signal Processing

Background:

• Human-Machine Interfaces (HMIs) enable natural user-machine interaction but require frequent re-calibration.
• Dynamic environments cause data drift, leading to HMI abandonment, an issue under-explored in Ultrasound-based (US-based) HMIs.
• Existing re-calibration methods often require labeled data, which is impractical for continuous adaptation.

Purpose of the Study:

• Investigate Unsupervised Domain Adaptation (UDA) algorithms for re-calibrating US-based HMIs during within-day sessions.
• Propose a novel CNN-based architecture for simultaneous wrist rotation and finger gesture prediction.
• Evaluate the effectiveness of UDA in improving HMI performance without labeled data.

Main Methods:

• Developed a Convolutional Neural Network (CNN) architecture for predicting wrist rotation angle and finger gestures.
• Implemented and evaluated various Unsupervised Domain Adaptation (UDA) algorithms, including Domain-Adversarial training (DANN).
• Conducted experiments using unlabeled data for re-calibration within dynamic environments.

Main Results:

• The proposed CNN architecture achieved state-of-the-art performance with 87.92% fewer trainable parameters.
• Domain-Adversarial training (DANN), with optimal initialization, demonstrated an average 24.99% increase in classification accuracy compared to no re-calibration.
• Performance enhancements from UDA were contingent on the consistency of the experimental setup and UDA configuration.

Conclusions:

• UDA techniques show promise for adapting US-based HMIs in dynamic environments without requiring labeled data.
• The proposed CNN architecture offers an efficient solution for gesture and rotation prediction.
• Careful consideration of experimental setup and UDA configuration is crucial for realizing significant performance gains.