Towards Black-Box Explainability With Gaussian Discriminant Knowledge Distillation

Anselm Haselhoff, Jan Kronenberger, Fabian Küppers, Jonas SchneiderWorkshops @ CVPR 2021 · DOI

Introduction

Modern deep learning models are often "black boxes"—they provide highly accurate predictions but offer no insight into why they made a specific choice. While existing methods like saliency maps highlight important pixels, they lack semantic meaning. This paper introduces a novel "student-teacher" framework designed to explain black-box AI decisions using human-understandable concepts like "arms" or "legs." By training an explainable generative student model to mimic a complex teacher, researchers can quantify the exact contribution of specific semantic concepts to any given prediction.

The Framework: Gaussian Discriminant Knowledge Distillation

The proposed approach utilizes a generative Knowledge Distillation (KD) framework consisting of three core components:

  1. Concept Model: A pre-trained ResNet50 classifier that identifies high-level semantic concepts (e.g., "torso," "hand," "head") within an image.
  2. Concept Probability Density Encoder (CPDE): The "student" model's core, which predicts the mean and covariance of concept distributions to match the "teacher" model's output.
  3. Gaussian Discriminant Decoder (GDD): This component uses Gaussian Discriminant Analysis (GDA) to convert concept scores into a final class prediction.
Overview of the proposed student-teacher framework showing the interaction between the teacher, concept model, and the student encoder-decoder architecture.

The system is optimized using a dual-objective loss function that balances Fidelity (the ability to mimic the teacher) and Explainability (ensuring student distributions remain grounded in real-world concept data).

Quantitative Semantics via Log Likelihood Ratio

A key innovation of this research is the use of the Log Likelihood Ratio (LLR) to provide quantitative explanations. Because the GDD uses a closed-form probabilistic approach based on Bayes' theorem, it allows for the calculation of a concrete measure for each concept's contribution to the final decision.

Qualitative example showing how concept presence and contribution (LLR) explain low model confidence in a person detection task.

As shown in the figure above, when a model has low confidence in a "person" detection (for instance, due to occlusion in a refrigerator), the LLR graph explicitly identifies which missing parts—such as the legs—caused the drop in confidence. This provides a much richer, semantically meaningful explanation than a standard heatmap.

Key Results and Insights

The researchers evaluated the framework using the DensePose dataset, focusing on human body part detection. Key findings include:

  • High Fidelity: The student model achieved a 0.81 Pearson correlation with the teacher's confidence scores, significantly outperforming standard GDA baselines (0.69).
  • Accuracy: In person detection tasks, the student model reached 94.48% accuracy, nearly matching the teacher's 95.83%, proving it successfully learned the teacher's underlying logic.
  • Handling Bimodal Data: The generative student model successfully captured bimodal distributions (e.g., for "foot" and "leg"), which improved its ability to mimic the teacher compared to simpler linear models.

Conclusion

This framework represents a significant step toward post-hoc explainability for black-box models. By grounding explanations in semantic concepts and providing quantitative contribution scores, it offers a path toward safer and more transparent AI in critical fields like autonomous driving and medical diagnostics. Future work will explore methods to discover concepts automatically and reduce the computational overhead of training separate student models.

Summary updated

BibTeX
@inproceedings{haselhoff2021towards,
  author = {Anselm Haselhoff and Jan Kronenberger and Fabian Küppers and Jonas Schneider},
  title = {{Towards Black-Box Explainability With Gaussian Discriminant Knowledge Distillation}},
  booktitle = {Workshops @ CVPR},
  year = {2021},
  doi = {10.1109/cvprw53098.2021.00011},
}
Figure