Globally interpretable models are a promising approach for trustworthy AI in safetycritical domains. Alongside global explanations, detailed local explanations are a crucial complement to effectively support human experts during inference. This work proposes the Calibrated Hierarchical QPM (CHiQPM) which offers uniquely comprehensive global and local interpretability, paving the way for human-AI complementarity. CHiQPM achieves superior global interpretability by contrastively explaining the majority of classes and offers novel hierarchical explanations that are more similar to how humans reason and can be traversed to offer a built-in interpretable Conformal prediction (CP) method. Our comprehensive evaluation shows that CHiQPM achieves state-of-the-art accuracy as a point predictor, maintaining 99% accuracy of non-interpretable models. This demonstrates a substantial improvement, where interpretability is incorporated without sacrificing overall accuracy. Furthermore, its calibrated set prediction is competitively efficient to other CP methods, while providing interpretable predictions of coherent sets along its hierarchical explanation.

Although theteacher haslearned rich and powerful representations, it also contains unignorable bias knowledge which is usually induced by the context prior (e.g., background) in the training data.

The existing distillation approaches mainly focus on using different criteria to align the sample representations learned by the student and the teacher, while they fail to transfer the class representations. Good class representations can benefit the sample representation learning by shaping the sample representation distribution. On the other hand, the existing approaches enforce the student to fully imitate the teacher while ignoring the fact that the teacher is typically not perfect. Although the teacher has learned rich and powerful representations, it also contains unignorable bias knowledge which is usually induced by the context prior (e.g., background) in the training data. To address these two issues, in this paper, we propose comprehensive, interventional distillation (CID) that captures both sample and class representations from the teacher while removing the bias with causal intervention. Different from the existing literature that uses the softened logits of the teacher as the training targets, CID considers the softened logits as the context information of an image, which is further used to remove the biased knowledge based on causal inference. Keeping the good representations while removing the bad bias enables CID to have a better generalization ability on test data and a better transferability across different datasets against the existing state-of-the-art approaches, which is demonstrated by extensive experiments on several benchmark datasets.

Globally interpretable models are a promising approach for trustworthy AI in safety-critical domains. Alongside global explanations, detailed local explanations are a crucial complement to effectively support human experts during inference. This work proposes the Calibrated Hierarchical QPM (CHiQPM) which offers uniquely comprehensive global and local interpretability, paving the way for human-AI complementarity. CHiQPM achieves superior global interpretability by contrastively explaining the majority of classes and offers novel hierarchical explanations that are more similar to how humans reason and can be traversed to offer a built-in interpretable Conformal prediction (CP) method. Our comprehensive evaluation shows that CHiQPM achieves state-of-the-art accuracy as a point predictor, maintaining 99% accuracy of non-interpretable models. This demonstrates a substantial improvement, where interpretability is incorporated without sacrificing overall accuracy. Furthermore, its calibrated set prediction is competitively efficient to other CP methods, while providing interpretable predictions of coherent sets along its hierarchical explanation.

We do this for all combinations of blocks and tokens. 1 2 Class representations in image tokens across the hierarchy Asterisks indicate a significant difference between both types of tokens. We additionally conducted an analysis comparing the class similarity change rate of class-and context-labeled tokens in self-attention layers. Figure 17: Agreement rate difference between correctly classified vs. misclassified samples. Figure 18: Percentage of instances where the layer's final predictions match any of the top-5 predictions of the most activated memories. AUC is better, while in the positive perturbation experiments (POS) a lower AUC is better.

The confusion matrix is a standard tool for evaluating classifiers by providing insights into class-level errors. In heterogeneous settings, its values are shaped by two main factors: class similarity -- how easily the model confuses two classes -- and distribution bias, arising from skewed distributions in the training and test sets. However, confusion matrix values reflect a mix of both factors, making it difficult to disentangle their individual contributions. To address this, we introduce bistochastic normalization using Iterative Proportional Fitting, a generalization of row and column normalization. Unlike standard normalizations, this method recovers the underlying structure of class similarity. By disentangling error sources, it enables more accurate diagnosis of model behavior and supports more targeted improvements. We also show a correspondence between confusion matrix normalizations and the model's internal class representations. Both standard and bistochastic normalizations can be interpreted geometrically in this space, offering a deeper understanding of what normalization reveals about a classifier.

On the other hand, as the teacher has learned rich and powerful representations, the existing approaches enforce the student to fully mimic the behavior of the teacher.