Magnetoencephalography Decoding Transfer Approach: From Deep Learning Models to Intrinsically Interpretable Models.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Deep learning models excel at decoding neuroelectrophysiological signals like Magnetoencephalography (MEG) but lack interpretability.
• This interpretability gap hinders reliability and ethical application in real-world scenarios.
• Intrinsically interpretable models offer transparency but often compromise predictive accuracy.

Purpose of the Study:

• To develop a method combining the high accuracy of deep learning with the interpretability of simpler models for MEG decoding.
• To enable the transformation of complex deep models into accurate and intrinsically interpretable models.

Main Methods:

• Pioneered an MEG transfer approach using feature attribution-based knowledge distillation.
• Introduced post-hoc feature knowledge, specifically feature attribution maps, into knowledge distillation for the first time.
• Guided intrinsically interpretable models (student) to assimilate knowledge from deep learning models (teacher).

Main Results:

• The proposed approach significantly outperformed benchmark knowledge distillation algorithms.
• Achieved up to an 8.28% improvement in prediction accuracy for Soft Decision Trees.
• Resulting interpretable models demonstrated performance comparable or superior to deep teacher models.

Conclusions:

• The feature attribution-based knowledge distillation effectively transfers MEG decoding information, yielding interpretable yet highly accurate models.
• This model-agnostic approach offers broad applicability for enhancing the trustworthiness of neuroimaging data analysis.
• The method bridges the gap between predictive performance and interpretability in complex neuroscientific machine learning applications.