🤖 AI Summary
Weak interpretability of vision transformers (e.g., ViT, Swin Transformer) and the difficulty of existing attention visualization methods in precisely localizing target objects under local window architectures pose significant challenges. To address this, we propose, for the first time, a gradient-free explanation method grounded in token-level statistical features—namely, mean and variance—extracted from LayerNorm operations. Our approach models per-layer token distribution characteristics and fuses multi-head attention maps to achieve structural adaptivity for sliding-window models. Evaluated on ImageNet-1K, our method achieves an average 23.6% improvement in Top-1 Intersection-over-Union over baseline methods (e.g., Chefer et al.), while attaining strong alignment with human annotations (Spearman’s ρ = 0.81). This demonstrates substantial gains in both localization accuracy and discriminative consistency.
📝 Abstract
Transformer becomes more popular in the vision domain in recent years so there is a need for finding an effective way to interpret the Transformer model by visualizing it. In recent work, Chefer et al. can visualize the Transformer on vision and multi-modal tasks effectively by combining attention layers to show the importance of each image patch. However, when applying to other variants of Transformer such as the Swin Transformer, this method can not focus on the predicted object. Our method, by considering the statistics of tokens in layer normalization layers, shows a great ability to interpret the explainability of Swin Transformer and ViT.
