![[2502.12108] Using the Path of Least Resistance to Explain Deep Networks](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2502.12108] Using the Path of Least Resistance to Explain Deep Networks
Summary
The paper introduces Geodesic Integrated Gradients (GIG), a new method for attributing importance scores in deep networks, addressing flaws in existing methods like Integrated Gradients (IG) by utilizing geodesic paths for more accurate attributions.
Why It Matters
Understanding and improving attribution methods in deep learning is crucial for model interpretability and trustworthiness. GIG proposes a significant enhancement over traditional methods, potentially leading to better insights in AI applications across various fields, including computer vision and AI safety.
Key Takeaways
- GIG offers a more accurate attribution method by using geodesic paths instead of straight lines.
- The introduction of the No-Cancellation Completeness axiom strengthens the theoretical foundation of attributions.
- Empirical results show GIG outperforms existing methods like Integrated Gradients on benchmark datasets.
Computer Science > Machine Learning arXiv:2502.12108 (cs) [Submitted on 17 Feb 2025 (v1), last revised 24 Feb 2026 (this version, v2)] Title:Using the Path of Least Resistance to Explain Deep Networks Authors:Sina Salek, Joseph Enguehard View a PDF of the paper titled Using the Path of Least Resistance to Explain Deep Networks, by Sina Salek and 1 other authors View PDF HTML (experimental) Abstract:Integrated Gradients (IG), a widely used axiomatic path-based attribution method, assigns importance scores to input features by integrating model gradients along a straight path from a baseline to the input. While effective in some cases, we show that straight paths can lead to flawed attributions. In this paper, we identify the cause of these misattributions and propose an alternative approach that equips the input space with a model-induced Riemannian metric (derived from the explained model's Jacobian) and computes attributions by integrating gradients along geodesics under this metric. We call this method Geodesic Integrated Gradients (GIG). To approximate geodesic paths, we introduce two techniques: a k-Nearest Neighbours-based approach for smaller models and a Stochastic Variational Inference-based method for larger ones. Additionally, we propose a new axiom, No-Cancellation Completeness (NCC), which strengthens completeness by ruling out feature-wise cancellation. We prove that, for path-based attributions under the model-induced metric, NCC holds if and only if the inte...