CVPR Findings2026
Value bounds and Convergence Analysis for Averages of LRP attributions
Alexander Binder, Nastaran Takmil-Homayouni, Urun Dogan
Abstract
We analyze theoretical properties of Layer-wise relevance propagation (LRP)-type attribution methods. Our contributions add novel insights to the theory of LRP and robustifies LRP-γ in practice. (1) We derive for the first time bounds of value ranges for attribution maps for LRP-β and LRPγ. (2) We use the value range bounds from the previous step to establish convergence speeds for averages of LRP attribution maps originating from test time averaging and SmoothGrad. We experimentally validate that these averages converge faster than for the gradient. (3) We establish a connection between value range stability and attribution map faithfulness: We use the newly found condition in Theorem 4 to dynamically adjust the γ parameter in LRP-γ which results in substantially improved faithfulness of the attribution maps to the model predictions.
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
Lineage
Builds on12
- EfficientNetV2: Smaller Models and Faster TrainingMingxing Tan, Quoc V. LeICML 2021 · 4,239 citations
- Swin Transformer V2: Scaling Up Capacity and ResolutionZe Liu, Han Hu, Yutong Lin, Zhuliang Yao et al.CVPR 2022 · 2,138 citations
- Better Aggregation in Test-Time AugmentationDivya Shanmugam, Davis W. Blalock, Guha Balakrishnan, John V. GuttagICCV 2021 · 205 citations
- A Consistent and Efficient Evaluation Strategy for Attribution MethodsYao Rong, Tobias Leemann, Vadim Borisov, Gjergji Kasneci et al.ICML 2022 · 138 citations
- Which Explanation Should I Choose? A Function Approximation Perspective to Characterizing Post Hoc ExplanationsTessa Han, Suraj Srinivas, Himabindu LakkarajuNeurIPS 2022 · 126 citations
Related papers
- When Explanations Lie: Why Many Modified BP Attributions FailLeon Sixt, Maximilian Granz, Tim LandgrafICML 2020 · 147 citations
- XAI for Transformers: Better Explanations through Conservative PropagationAmeen Ali, Thomas Schnake, Oliver Eberle, Grégoire Montavon et al.ICML 2022 · 144 citations
- Interpreting Super-Resolution Networks With Local Attribution MapsJinjin Gu, Chao DongCVPR 2021
- Distilled Gradient Aggregation: Purify Features for Input Attribution in the Deep Neural NetworkGiyoung Jeon, Haedong Jeong, Jaesik ChoiNeurIPS 2022 · 11 citations
- Empowering CAM-Based Methods with Capability to Generate Fine-Grained and High-Faithfulness ExplanationsChangqing Qiu, Fusheng Jin, Yining ZhangAAAI 2024 · 11 citations
