MARGIN: Uncovering Deep Neural Networks Using Graph Signal Analysis.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Interpretability is vital for trust in opaque machine learning models.
• Existing methods address specific interpretability tasks like sample identification or prediction explanation.
• A unified approach is needed to handle diverse interpretability challenges.

Purpose of the Study:

• To introduce MARGIN, a general method for machine learning interpretability.
• To demonstrate that many interpretability tasks are variants of identifying relative prediction changes.
• To provide a unified framework for addressing a wide range of interpretability problems.

Main Methods:

• MARGiN (Measuring And Refining Graph-based INterpretability) leverages graph signal analysis.
• It identifies influential nodes in task-specific graphs that best describe a function.
• The approach defines graphs and functions tailored to specific interpretability challenges.

Main Results:

• MARGiN successfully addresses a broad spectrum of interpretability tasks.
• The method demonstrates superior performance compared to existing techniques.
• It provides a unified solution for problems previously treated separately.

Conclusions:

• MARGiN offers a simple yet generalizable approach to machine learning interpretability.
• The method's foundation in relative prediction change unifies diverse tasks.
• This work advances the field by providing a more cohesive and effective interpretability solution.