AlphaGrad: Normalized Gradient Descent for Adaptive Multi-Loss Functions in EEG-Based Motor Imagery Classification.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Neuroscience

Background:

• Multi-task learning (MTL) models are increasingly used for electroencephalography (EEG) classification.
• Optimizing MTL models often involves balancing multiple loss functions, which can be challenging due to differing metric scales.
• Existing methods for loss blending in motor imagery (MI)-EEG classification lack adaptability.

Purpose of the Study:

• To introduce AlphaGrad, a novel adaptive loss blending strategy for optimizing MTL models in MI-EEG classification.
• To address the challenge of automatically adjusting multi-loss functions with differing metric scales.
• To enhance classification accuracy and training stability in MI-EEG analysis.

Main Methods:

• Proposed AlphaGrad, an adaptive loss blending strategy for MTL.
• Evaluated AlphaGrad on two state-of-the-art MTL neural networks (MIN2Net, FBMSNet) across four benchmark datasets.
• Utilized gradient trajectory visualizations to analyze training stability and local minima avoidance.

Main Results:

• AlphaGrad consistently outperformed existing strategies (AdaMT, GradApprox, fixed-weight baselines) in classification accuracy and training stability.
• Achieved over 10% accuracy improvement on subject-independent MI tasks compared to static weighting.
• Demonstrated robust adaptability across various EEG paradigms (SSVEP, ERP), indicating broad applicability to brain-computer interface (BCI) systems.

Conclusions:

• AlphaGrad is an effective general-purpose solution for adaptive multi-loss optimization in biomedical time-series learning.
• The method enhances performance and stability in MI-EEG classification.
• AlphaGrad shows significant promise for advancing BCI systems through improved optimization techniques.