AutoTransfer: Subject Transfer Learning with Censored Representations on Biosignals Data.

Area of Science:

• Machine Learning
• Biomedical Signal Processing
• Computational Neuroscience

Background:

• Subject transfer learning is crucial for analyzing biomedical data where subject variability is high.
• Existing methods often struggle with generalizing across different subjects due to domain shifts.
• Developing robust regularization techniques is essential for improving cross-subject generalization.

Purpose of the Study:

• To propose a novel regularization framework for subject transfer learning.
• To introduce methods for measuring and enforcing independence between latent representations and subject labels.
• To develop a practical, automated strategy for applying these regularization techniques.

Main Methods:

• Developed a regularization framework minimizing classification loss with an independence penalty.
• Introduced three notions of independence using mutual information or divergence.
• Implemented estimation algorithms using analytic methods and neural critic functions.
• Proposed the 'Auto Transfer' hands-off strategy for applying regularization schemes.

Main Results:

• Demonstrated improved performance in subject transfer learning across EEG, EMG, and ECoG datasets.
• Showcased the effectiveness of different independence-based regularization strategies.
• Validated the 'Auto Transfer' framework's utility on challenging, real-world biomedical data.

Conclusions:

• The proposed regularization framework significantly enhances subject transfer learning.
• Independence-based penalties are effective in improving cross-subject generalization.
• The 'Auto Transfer' strategy provides a practical approach for applying these methods to new datasets.