Attribution Upsampling should Redistribute, Not Interpolate

Current explanation tools are using the wrong technique: Standard methods for enlarging saliency maps (bilinear and bicubic interpolation) were designed for photographs, not explanations. These techniques create artificial "hot spots" that don't actually reflect what the AI model is focusing on, misleading users about how the model makes decisions.

The problem is fundamental, not minor: The paper proves that interpolation-based upsampling violates three of four essential principles for faithful AI explanations. This means the issue isn't a small bug to fix—it's a structural flaw in how current explanation methods work.

A better approach treats explanations like redistribution, not resizing: Rather than stretching importance scores like you would stretch an image, the proposed USU method redistributes importance based on the model's actual semantic understanding. Testing on ImageNet, CIFAR-10, and CUB-200 datasets shows this produces explanations that are more faithful to what the model actually learned.

Improved explanations lead to more trustworthy AI: By fixing how importance is displayed in visual explanations, users and auditors can more accurately understand model reasoning. This is critical for high-stakes applications where people need to trust that AI explanations genuinely reflect model behavior rather than artifacts of the visualization process.