Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Manifold-based Shapley for SAR Recognization Network Explanation

Xuran Hu, Mingzhe Zhu, Yuanjing Liu, Zhenpeng Feng, LJubisa Stankovic

TL;DR

This work tackles SAR explainability by addressing the independence assumption of Shapley values in high-dimensional features. It introduces a manifold-based Shapley using StyleGAN2 to map high-dimensional SAR features to a low-dimensional manifold and reconstructs back via Image2StyleGAN, enabling credible attributions on the manifold. To overcome mapping and dimensionality challenges, it proposes a gradient-based Shapley mapping to the original space and fuses manifold and traditional Shapley through a fusion coefficient $\alpha$, yielding Fusion-Shap. On the MSTAR SAR dataset, Fusion-Shap improves fidelity, sensitivity, and axiomatic validity over existing methods, offering more robust and interpretable explanations for SAR recognition models.

Abstract

Explainable artificial intelligence (XAI) holds immense significance in enhancing the deep neural network's transparency and credibility, particularly in some risky and high-cost scenarios, like synthetic aperture radar (SAR). Shapley is a game-based explanation technique with robust mathematical foundations. However, Shapley assumes that model's features are independent, rendering Shapley explanation invalid for high dimensional models. This study introduces a manifold-based Shapley method by projecting high-dimensional features into low-dimensional manifold features and subsequently obtaining Fusion-Shap, which aims at (1) addressing the issue of erroneous explanations encountered by traditional Shap; (2) resolving the challenge of interpretability that traditional Shap faces in complex scenarios.

Manifold-based Shapley for SAR Recognization Network Explanation

TL;DR

This work tackles SAR explainability by addressing the independence assumption of Shapley values in high-dimensional features. It introduces a manifold-based Shapley using StyleGAN2 to map high-dimensional SAR features to a low-dimensional manifold and reconstructs back via Image2StyleGAN, enabling credible attributions on the manifold. To overcome mapping and dimensionality challenges, it proposes a gradient-based Shapley mapping to the original space and fuses manifold and traditional Shapley through a fusion coefficient , yielding Fusion-Shap. On the MSTAR SAR dataset, Fusion-Shap improves fidelity, sensitivity, and axiomatic validity over existing methods, offering more robust and interpretable explanations for SAR recognition models.

Abstract

Explainable artificial intelligence (XAI) holds immense significance in enhancing the deep neural network's transparency and credibility, particularly in some risky and high-cost scenarios, like synthetic aperture radar (SAR). Shapley is a game-based explanation technique with robust mathematical foundations. However, Shapley assumes that model's features are independent, rendering Shapley explanation invalid for high dimensional models. This study introduces a manifold-based Shapley method by projecting high-dimensional features into low-dimensional manifold features and subsequently obtaining Fusion-Shap, which aims at (1) addressing the issue of erroneous explanations encountered by traditional Shap; (2) resolving the challenge of interpretability that traditional Shap faces in complex scenarios.
Paper Structure (13 sections, 13 equations, 4 figures, 1 table)

This paper contains 13 sections, 13 equations, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methodology
  3. Calculation of data manifold
  4. Manifold-based Shapley
  5. Shapley mapping
  6. Shapley Fusion
  7. Experimental Results
  8. Experiment Settings
  9. Visualization and Subjective Evaluation
  10. Axiomatic Validation
  11. Explanation Sensitivity and Infidelity
  12. Manifold Dimension and Explanation
  13. Conclusions

Figures (4)

  • Figure 1: UMAP mcinnes2018umap-software visualization of low-dimensional manifolds and high-dimensional features.
  • Figure 2: Left: Implementation of Fusion-shap; Right: The black-box model under interpretation:$F$. StyleGAN generators($G$) and Image2StyleGAN($R$), which enable the transformation of high-dimensional feature and low-dimensional manifolds.
  • Figure 3: Results visualization. Columns from left to right: Origin image, Grad CAM, LRP, IG, SG, SHAP and F-SHAP.
  • Figure 4: Infidelity and sensitivity vary across different dimensions of manifold.