Explainable AI for ML jet taggers using expert variables and layerwise relevance propagation

Expert knowledge can make AI systems more interpretable and performant: By incorporating domain expert variables (XAUG variables) alongside raw data inputs, researchers improved classifier efficiency by 30-40% while simultaneously making the neural network's decision-making process more understandable to humans.

Explainability techniques can identify which features matter most: Layerwise relevance propagation (LRP) allowed researchers to rank which input features the AI system actually relied on for decisions, enabling them to identify a reduced set of features that capture the network's performance—a practical step toward simpler, more auditable AI systems.

Combining human expertise with AI can reveal what models are actually learning: When comparing networks with and without expert variables, researchers found that XAUG variables could "capture the behavior of the classifier completely," suggesting this approach can validate whether AI systems are learning the right patterns or relying on spurious correlations.

Uncertainty quantification matters for responsible deployment: The paper presents methods to quantify numerical uncertainties in neural network training, addressing a key concern for responsible AI: understanding not just what a model predicts, but how confident we should be in those predictions.