Optimizing Explanations by Network Canonization and Hyperparameter Search

Model canonization improves XAI reliability across architectures: The paper demonstrates that restructuring neural networks to remove problematic components (like BatchNorm layers) without changing their function makes explanation methods more consistent and trustworthy, addressing a fundamental issue where different implementations of the same model could produce different explanations.

Implementation-invariance matters for trust: The research highlights that many current XAI methods violate the "implementation-invariance axiom"—meaning the same model can produce different explanations depending on how it's technically implemented. This is a critical problem for deploying AI systems in high-stakes applications where consistency is essential.

Hyperparameter tuning significantly affects explanation quality: The paper shows that XAI methods have numerous potential parameterizations, and the choice of these parameters substantially impacts the explanations produced. This suggests that practitioners cannot simply apply XAI methods "out of the box" without careful optimization.

Canonization works for modern architectures: The authors successfully develop canonization techniques for currently popular deep learning models (EfficientNet, DenseNets, Relation Networks), not just older architectures, making their approach practically relevant for today's AI systems used in real applications.

Quantifiable evaluation framework is needed: The paper establishes a measurable evaluation framework for XAI methods across multiple tasks (image classification and Visual Question Answering), addressing the broader need for standardized, objective ways to assess whether explanations are actually reliable and useful.