Global Explanations for Image Classifiers (Student Abstract)

Bhavan K. Vasu, Prasad TadepalliAAAI 2023 · DOI

Abstract

Deep neural networks are often criticized as "black boxes" because their decision-making processes are opaque. While local explanation methods like saliency maps highlight important pixels in individual images, they fail to provide a "big picture" understanding of how a model classifies entire categories. This paper introduces a method for generating global, symbolic explanations by identifying Minimal Sufficient Explanations (MSXs) and using matrix decomposition to extract core features. The approach successfully identifies discriminative objects for scene categories like "Living Room" and "Dining Room," offering a path toward more interpretable and debuggable AI.

Introduction

Understanding the logic behind deep learning models is crucial for trust and safety. Current methods primarily focus on local explanations, which are often subjective and image-specific. For instance, a saliency map might highlight a window in a "Living Room" image, but it doesn't explain if the model considers windows essential for all living rooms. This research addresses the need for global explanations that summarize a model's logic across a dataset, moving from pixel-level highlights to human-understandable symbolic features.

Methodology

The proposed pipeline consists of three main stages:

  1. Local Explanations via Beam Search: The system identifies Minimal Sufficient Explanations (MSXs) for individual images. An MSX is the smallest set of objects that maintains at least 95% of the model's original prediction confidence. This is achieved by systematically blurring out objects using a beam search algorithm.
  2. Symbolic Representation: These local explanations are aggregated into a sparse matrix $G$, where rows represent object categories and columns represent individual local explanations.
  3. Global Explanations via SVD: Singular Value Decomposition (SVD) is applied to the matrix $G$. This low-rank approximation filters out noise and identifies the most significant objects that define a class globally.
Local explanations (MSXs) for a dining room scene.

Evaluation and Datasets

The method was evaluated using the ADE20K dataset, specifically focusing on the top 10 scene categories. A VGG19 model was trained on these categories, achieving 98% test accuracy. The researchers then applied their global explanation pipeline to understand what the VGG19 model had actually learned.

Original dining room scene from the dataset.

Key Insights and Results

The global explanations provided several critical insights:

  • Discriminative Features: By comparing the sets of objects for different classes, the researchers identified how the model distinguishes similar scenes. For example, while both living and dining rooms share "walls" and "floors," the model specifically looks for "sofas" and "rugs" to identify a Living Room, and "tables" and "chairs" for a Dining Room.
  • Complexity Reduction: The SVD-based approach significantly pruned the list of relevant objects, distilling complex scenes down to a core set of 4–5 essential items, which aligns closely with human intuition.
  • Model Debugging: The symbolic nature of these explanations allows developers to see if a model is relying on spurious correlations or if it has truly captured the essence of a category.

Conclusion and Future Work

This work demonstrates a viable path for moving from local pixel-based explanations to global symbolic logic. Future research will focus on integrating standardized object taxonomies to improve human interpretability and conducting user studies to measure the impact of these explanations on human trust. Additionally, the authors plan to explore more complex symbolic representations beyond simple presence/absence matrices.

Summary updated

BibTeX
@inproceedings{vasu2023global,
  author = {Bhavan K. Vasu and Prasad Tadepalli},
  title = {{Global Explanations for Image Classifiers (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.1609/aaai.v37i13.27036},
}
Figure