Knowledge distillation affects how small networks explain their image decisions
Transferring Visual Explanations: How Cross-Architecture Knowledge Distillation Affects Model Interpretability
Computer Vision and Pattern Recognition
Summary
Smaller neural networks are used when there isn't a lot of computing power, but they often lose some ability to explain how they make decisions. This paper studies if teaching a small model by copying a large one also passes on the way the large model points to important image parts. The authors find that copying helps preserve rough attention to objects but not fine details in explanations. They also show that the student's own design limits how well detailed explanations can be learned from the teacher.
What this means in practice
- •For autonomous vehicle engineers: Build smaller neural networks that maintain rough explanation of visual decisions for safety-critical system validation.
- •For medical device developers: Develop compact models that preserve coarse localization explanations to improve trust in diagnostic AI under resource constraints.
Tested on one dataset.
Authors
Aleks Czufarow, Ihor Babin
Abstract
Deploying efficient neural networks is essential in resource-constrained environments, yet compact models often sacrifice interpretability - a critical in safety-critical domains such as autonomous driving and medicine. This study investigates whether Knowledge Distillation transfers the spatial feature attribution of a large teacher network to a compact student. To assess the influence of the KD scheme on interpretability, we distill a ResNet-152 teacher into a ResNet-34 student on ImageNet-1K across five configurations by systematically varying the distillation temperature and soft-label loss weight. Models are evaluated on top-1 accuracy, along with two interpretability metrics: Relevance Mass Accuracy and Relevance Rank Accuracy. These metrics are computed via Grad-CAM heatmaps benchmarked against ground-truth object masks. Our results show that top-1 accuracy ranges from 71.6% to 74.0%. For Grad-CAM, RMA ranges from 7.7% to 9.7% and RRA from 7.3% to 10.1%; for Guided Grad-CAM, RMA ranges from 16.1% to 18.6% and RRA from 15.9% to 21.5%. Interpretability proves far more sensitive to the soft-label weight than to the temperature: keeping the student anchored to hard labels preserves both accuracy and coarse localization, whereas weighting the teacher heavily degrades both. Fine-grained attribution, however, fell below the undistilled baseline in every configuration tested, indicating that logit distillation transmits where a model attends more readily than the pixel-level structure of that attention. We evaluate 12 cross-architecture combinations of convolutional and transformer-based models, revealing that the inheritance of fine-grained spatial reasoning is fundamentally bottlenecked by the student's intrinsic structural biases. To our knowledge, this is the first application of this interpretability-aware evaluation framework - previously used for neural network pruning - to KD.