AGOP as Explanation: From Feature Learning to Per-Sample Attribution in Image Classifiers

AGOP-Weighted offers a practical improvement over existing explanation methods: The new attribution approach achieved 44% higher accuracy (measured by mIoU) compared to Integrated Gradients on linear classification tasks, suggesting it can provide more reliable explanations of what image features a classifier is using to make decisions.

Training data patterns can improve individual prediction explanations: By incorporating information about which pixels are consistently important across the training distribution, AGOP-Weighted suppresses noise in explanations while amplifying genuinely important features—a combination the authors note hasn't been used in prior attribution methods.

Computational efficiency enables practical deployment: AGOP-Global requires zero inference cost (only a disk lookup) and AGOP-Weighted needs just a single gradient computation, making these methods feasible for real-world applications where computational resources are limited, unlike more expensive alternatives like Integrated Gradients.

Rigorous benchmarking with pixel-level ground truth strengthens credibility: The evaluation used two benchmarks with known correct answers (XAI-TRIS and CLEVR-XAI), allowing direct measurement of explanation accuracy rather than relying on proxy metrics—an important standard for validating that explanation methods actually work as claimed.

Theoretical foundation connects feature learning to explanations: The work builds on established neural network theory (the Neural Feature Ansatz) to justify why AGOP should work for explanations, providing principled reasoning rather than purely empirical justification for the approach.