Fig. 5: An example of the shortcut hull learning.
From: Mitigating data bias and ensuring reliable evaluation of AI models with shortcut hull learning
For each data class, we select a representative image as an example. We then showcase the features of this image as processed by models with different inductive biases, including ResNet-5026, ViT-B/1641, RepVGG-A257, Swin-T58, and PViG-S59. These features are projected onto the input image using the HiResCAM56 method. After thresholding, we obtain features corresponding to the pixel positions of the image. The common features across different models are then computed using Eq. (8). Beneath each feature, the red numbers indicate the total count of pixel position features.
