Why Does It Look There? Structured Explanations for Image Classification

Existing AI explanations often rely on external AI systems, which can distort the truth: The paper identifies a key problem with current explainability methods—they frequently use auxiliary AI models like GPT or CLIP to describe how image classifiers work, which compromises "faithfulness to the original models." I2X addresses this by explaining models directly without intermediaries.

The framework reveals how models make decisions, not just where they look: Rather than just showing saliency maps (highlighted regions), I2X provides "structured explanations" that track both "intra- and inter-class decision making during training," offering a more complete picture of model reasoning for different image categories.

Understanding model development can improve performance: The paper demonstrates a practical benefit beyond transparency—by identifying "uncertain prototypes" that I2X reveals, researchers can strategically fine-tune models to improve accuracy, suggesting that explainability and performance improvement can go hand-in-hand.

The approach works across different models and datasets: Experiments on MNIST and CIFAR10 show I2X's effectiveness "across different model architectures and datasets," indicating the method has broad applicability rather than being limited to specific scenarios.

Transparency doesn't require sacrificing model fidelity: I2X shows it's possible to create trustworthy explanations that remain faithful to how the original model actually works, addressing a core tension in responsible AI between interpretability and accuracy.