Abstract: While convolution neural networks (CNNs) and vision transformers (ViTs) dominate visual representation learning, the growing model depth causes difficulty for interpretability. Although ...