Overview
GradSkip is a new relevance propagation method specifically designed for Vision Transformers (ViTs). It addresses limitations in existing interpretation techniques by considering architectural features such as the varying importance of attention heads and the role of residual connections. The method incorporates adaptive head weighting and skip-aware propagation to refine how relevance is attributed within ViTs.
Research Context
Vision Transformers (ViTs) present challenges for interpretation because current relevance propagation and attention flow methods do not fully account for certain architectural characteristics. Prior approaches typically operate under assumptions that do not align with ViT design. Specifically, they often assume uniform importance across attention heads and model skip connections as straightforward identity paths. These assumptions can lead to inaccuracies in relevance attribution, making it difficult to understand which parts of an input image contribute most to a ViT's decision.
The research establishes that the uneven importance of attention heads and the functionality of residual connections are critical architectural features that need to be considered for accurate interpretation of ViTs. The goal was to develop a method that more accurately reflects how relevance propagates through these complex structures.
Approach
The proposed method, GradSkip, is a relevance propagation technique for ViTs. Its design is predicated on two core principles: adaptive head weighting and skip-aware propagation. Unlike previous methods that assume uniform importance, GradSkip models the distinct importance levels of individual attention heads. Furthermore, it dynamically distributes relevance between the attention and residual paths, moving beyond the simple identity path model for skip connections. This dynamic distribution mechanism allows GradSkip to adaptively account for the contribution of both the transformed feature and the original feature element in the residual connection. The method aims to provide a more nuanced and accurate understanding of how ViTs arrive at their outputs by explicitly incorporating these architectural considerations into the relevance propagation process.
Findings
Experimental evaluations of GradSkip were conducted on two distinct datasets: ImageNet1K and BloodMNIST. The results demonstrate that GradSkip achieves state-of-the-art faithfulness when compared to existing approaches for ViT interpretation. A notable aspect of its performance is its computational efficiency, requiring over 14 times fewer GFLOPs than the best-performing interpretative methods currently available. This indicates a significant reduction in computational resources needed for relevance attribution.
Further assessments were carried out using transformer-based segmentation tasks. In these evaluations, GradSkip exhibited improved localization capabilities. The method also showed enhanced alignment with ground-truth regions, suggesting a more precise identification of salient features contributing to the model's output in segmentation contexts.