Introduction
Deep neural networks (DNNs) have achieved state-of-the-art performance in various computer vision tasks, yet they remain largely "black boxes." Understanding the rationale behind a specific prediction is crucial for high-stakes applications like healthcare or autonomous driving. Existing methods often struggle to provide precise, non-blurry explanations that accurately reflect the local decision boundary of the model. This paper introduces a novel "re-label distillation" framework to generate high-fidelity saliency maps by probing the model's behavior around a specific input.
Methodology: Re-Label Distillation
The proposed approach generates post-hoc explanations through a three-stage pipeline designed to capture the local characteristics of a DNN's decision-making process.
1. Self-supervised Image Synthesis
The process begins with a pre-trained Variational Autoencoder (VAE). For a target input image, the VAE projects it into a latent space. By adding random noise to this latent representation, the system generates a diverse set of synthetic images. These images are semantically similar to the original but contain subtle variations that help explore the surrounding feature space.
2. Re-labeling and Boundary Mapping
The synthetic images are passed through the "teacher" DNN. The resulting predictions are used to re-label the synthetic set into two categories: those that maintain the original classification and those that trigger a class shift. This binary classification task effectively maps the local decision boundary of the complex, non-linear teacher model.
3. Linear Distillation
A simple, two-class linear student model is then trained on these re-labeled synthetic images. By minimizing a combined distillation and classification loss, the student model learns to approximate the teacher's local behavior. Because the student model is linear, its weights directly represent the importance of each pixel, which are then visualized as a saliency map.
Key Insights
The core innovation lies in "probing the boundary." Rather than just looking at what features are present, the method identifies which features, when altered, cause the model to change its prediction. This focus on the decision threshold allows the method to extract "boundary knowledge" that traditional gradient-based methods often miss. Furthermore, the use of a linear model for local approximation ensures that the resulting explanations are mathematically grounded and easily interpretable.
Experimental Results
The method was evaluated on ImageNet using ResNet50 and VGG16 architectures. Quantitative benchmarks using "Deletion" and "Insertion" metrics demonstrated superior performance. For instance, on ResNet50, the method achieved a Deletion score of 0.0627, significantly outperforming Grad-CAM (0.1421).
Qualitatively, as shown in the figure above, the saliency maps produced by re-label distillation are notably more precise. While other methods often highlight broad regions or background noise, the proposed approach focuses tightly on discriminative features, such as the beak of a bird or the specific texture of a lemon.
Conclusion
Re-label distillation offers a robust way to interpret deep networks by distilling complex local decision boundaries into simple linear models. While it incurs higher computational costs due to the generation of synthetic images, the resulting clarity and accuracy of the saliency maps provide a significant step forward in making AI systems more transparent and trustworthy. Future work may focus on optimizing the synthesis process and extending the framework to non-image data domains.
Summary updated