Data-related Ablation for Reinforcing Deep Learning in Explaining Complex Phenomena.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Neuroscience

Background:

• Deep learning (DL) models excel at pattern recognition but suffer from a "black box" nature, hindering trust.
• Current validation methods focus on model architecture, overlooking potential data biases.
• Implicit trust in data can lead to misleading performance evaluations.

Purpose of the Study:

• To introduce a novel "data-related ablation" technique as a complement to traditional architectural ablation.
• To evaluate the reliability and generalizability of DL models by assessing their reliance on data characteristics versus true patterns.
• To improve trust and transparency in DL models, particularly in complex data domains.

Main Methods:

• Developed a data-related ablation framework to complement architectural ablation.
• Applied the framework to Electroencephalography (EEG) signals for Emotional Recognition (ER) and Motor Execution (ME) tasks.
• Assessed model performance by observing its behavior when process-irrelevant features were eliminated.

Main Results:

• High-accuracy DL models often rely heavily on process-irrelevant features.
• Models maintained performance even when crucial information was removed, indicating reliance on data quirks.
• Standard, data-independent evaluations can be deceptive regarding true learning.

Conclusions:

• Data-related ablation is crucial for distinguishing robust learning from reliance on incidental data characteristics.
• The proposed method enhances the reliability and generalizability of DL models.
• This approach is essential for fields using complex, potentially biased data, like EEG analysis.