Unsupervised Transfer Learning Approach With Adaptive Reweighting and Resampling Strategy for Inter-subject EOG-based...

Area of Science:

• Biomedical Engineering
• Machine Learning
• Human-Computer Interaction

Background:

• Electrooculogram (EOG)-based gaze estimation is widely researched but faces challenges due to significant inter-subject variability.
• This variability leads to performance degradation in practical gaze estimation applications.

Purpose of the Study:

• To propose an unsupervised transfer learning approach with an adaptive reweighting and resampling (ARR) strategy to address inter-subject variability in EOG-based gaze angle estimation.
• To quantify domain shifts and adapt source data for improved model training.

Main Methods:

• An unsupervised transfer learning approach utilizing an adaptive reweighting and resampling (ARR) strategy.
• Quantifying domain shifts by decomposing transformation matrices between multi-source and target domains.
• Reweighting and resampling source data based on weighted indicators derived from domain shifts.

Main Results:

• The ARR strategy significantly improved gaze estimation performance, reducing mean absolute error (MAE) by 7.0% and root mean square error (RMSE) by 6.3%.
• The proposed method outperformed prevailing transfer learning techniques like CORAL, GFK, JDA, TCA, and BDA.
• Data size was found to be more critical than data diversity for the ARR strategy's effectiveness.

Conclusions:

• The adaptive reweighting and resampling (ARR) strategy effectively mitigates inter-subject variability in EOG-based gaze angle estimation.
• The ARR strategy offers a robust solution for practical gaze estimation scenarios, demonstrating superior performance over existing methods.