Attribution via Distributional Paths for Information Revelation

Problem
Current feature attribution methods, particularly path-based techniques like Integrated Gradients, face limitations in their reliance on input-space trajectories. These methods lack control over the resolution of feature queries, leading to potential artifacts in attribution scores. This paper addresses the gap in the literature regarding the completeness and stability of attributions derived from input-space paths, proposing a new approach that operates in a distributional space rather than directly in input space. The work is presented as a preprint and has not yet undergone peer review.

Method
The authors propose Reveal-IG, which shifts the focus from pointwise perturbations in input space to a structured probe distribution around the input of interest. This method progressively reveals information about the input, allowing for a more nuanced attribution of changes in the model’s expected output. Reveal-IG maintains completeness concerning the expected model response while accommodating multiscale image probes and feature-wise uncertainty in tabular data. The authors provide synthetic diagnostics to demonstrate that Reveal-IG effectively mitigates path artifacts that are prevalent in traditional input-space methods. The architecture and specific loss functions used in the training of Reveal-IG are not disclosed, but the method is evaluated on standard benchmarks.

Results
Reveal-IG is evaluated on ImageNet classification and tabular regression tasks, where it demonstrates superior performance in generating stable, signed attributions. The method leads on metrics that incorporate attribution sign, indicating a more reliable interpretation of feature importance. While specific numerical results are not detailed in the abstract, the authors claim that Reveal-IG remains competitive with existing methods on other attribution metrics, suggesting a robust performance across various evaluation criteria.

Limitations
The authors acknowledge that while Reveal-IG improves upon traditional path-based methods, it may still be sensitive to the choice of probe distributions and the underlying model architecture. Additionally, the paper does not address the computational overhead that may arise from using structured probe distributions, which could impact scalability in large datasets or real-time applications. The lack of peer review also raises questions about the robustness of the findings.

Why it matters
The introduction of Reveal-IG has significant implications for the field of interpretable machine learning, particularly in enhancing the reliability of feature attribution methods. By addressing the limitations of input-space path methods, this work paves the way for more accurate and informative explanations of model predictions, which is crucial for applications in sensitive domains such as healthcare and finance. The findings contribute to ongoing discussions about the interpretability of AI systems, as highlighted in related literature (as published in arXiv cs.LG).