MFPP: Morphological Fragmental Perturbation Pyramid for Black-Box Model Explanations

Qing Yang, Xia Zhu, Jong-Kae Fwu, Yun Ye, Ganmei You, Yuan Zhu

Abstract

Deep Neural Networks (DNNs) are often criticized for being "black boxes," providing accurate predictions without interpretable reasoning. This paper introduces Morphological Fragmental Perturbation Pyramid (MFPP), a novel model-agnostic explanation method. By utilizing multi-scale morphological fragments (superpixels) instead of traditional rigid grids, MFPP generates high-resolution saliency maps that respect object boundaries. Experimental results on PASCAL VOC and MS-COCO benchmarks demonstrate that MFPP outperforms state-of-the-art methods in both accuracy and computational efficiency, offering a significant speedup of up to 10x while maintaining superior visual clarity.

Introduction

The lack of transparency in DNNs poses a significant barrier to their deployment in high-stakes fields such as autonomous driving and medical diagnostics. Existing Explainable AI (XAI) techniques generally fall into two categories: white-box and black-box methods. White-box methods like Grad-CAM are fast but architecture-dependent and often produce low-resolution heatmaps. Black-box methods, such as RISE or LIME, are model-agnostic but typically rely on rigid grids or single-scale perturbations, which often fail to capture complex object geometries or fine-grained details.

MFPP addresses these limitations by introducing a perturbation strategy that is both fine-grained and computationally efficient. By leveraging the natural edges and textures of an image through superpixel segmentation, the method provides more intuitive and precise explanations of model behavior.

Methods

The core innovation of MFPP lies in its hierarchical approach to image perturbation. Unlike methods that mask square blocks, MFPP follows a multi-step pipeline:

  1. Morphological Segmentation: The input image is decomposed into "superpixels" using the SLIC algorithm. These fragments naturally align with the edges and textures of the scene.
  2. Pyramid Structure: Inspired by Feature Pyramid Networks (FPN), MFPP segments the image at multiple scales (e.g., five layers ranging from 50 to 800 fragments). This allows the explainer to capture both global context and minute details simultaneously.
  3. Random Perturbation and Aggregation: Thousands of perturbed versions of the image are created by randomly masking these multi-scale fragments. These are fed into the black-box classifier, and the resulting prediction scores are used as weights to aggregate the masks into a final saliency map.
The MFPP workflow: from input image and multi-scale superpixel segmentation to weighted summation of perturbed outputs.

To further optimize performance, the authors introduced Fast-MFPP, which reduces the number of required masks from 20,000 to 4,000, achieving a massive speedup with negligible impact on accuracy.

Experimental Setup

The researchers evaluated MFPP against gold-standard computer vision benchmarks, including PASCAL VOC 2007 and MS-COCO 2014. Performance was measured using the "Pointing Game" evaluation, which checks if the maximum intensity point of a saliency map falls within the object's bounding box. The experiments utilized standard architectures, specifically VGG16 and ResNet50, to ensure broad applicability.

Key Results and Insights

MFPP demonstrates clear superiority over existing methods in three key areas:

  • Accuracy: On the PASCAL VOC 2007 Pointing Game using ResNet50, MFPP achieved 89.1% accuracy, surpassing the previous state-of-the-art Extremal Perturbation (EP) at 88.9% and significantly outperforming RISE (86.6%).
  • Efficiency: MFPP is approximately 2.2x faster than the EP method. Fast-MFPP pushes this further, operating over 10x faster than EP while maintaining competitive accuracy.
  • Visual Fidelity: Qualitatively, MFPP produces much "cleaner" saliency maps. While Grad-CAM often results in vague blobs, MFPP can tightly trace specific shapes, such as the silhouette of a baseball player or individual pieces of fruit.
Qualitative comparison of saliency maps across different XAI methods. MFPP consistently produces more precise, boundary-aware explanations.

Conclusion

MFPP represents a significant advancement in black-box model interpretability. By replacing rigid grids with morphological fragments and adopting a pyramid structure, the method achieves a better balance between resolution and speed than its predecessors. While the method's quality is somewhat dependent on the initial superpixel segmentation, its ability to provide high-fidelity, model-agnostic explanations makes it a powerful tool for understanding complex deep learning models. Future work may involve exploring adaptive sampling techniques to further reduce the computational overhead of black-box explanations.

Summary updated

BibTeX
@inproceedings{yang2020mfpp,
  author = {Qing Yang and Xia Zhu and Jong-Kae Fwu and Yun Ye and Ganmei You and Yuan Zhu},
  title = {{MFPP: Morphological Fragmental Perturbation Pyramid for Black-Box Model Explanations}},
  booktitle = {ICPR},
  year = {2020},
  doi = {10.1109/icpr48806.2021.9413046},
}
Figure