From Local to Global to Mechanistic: An iERF-Centered Unified Framework for Interpreting Vision Models

This research presents a unified framework for understanding how vision models (AI systems that analyze images) work by focusing on pointwise feature vectors (PFVs, individual numerical representations at specific locations in an image) and their instance-specific Effective Receptive Fields (iERFs, the regions of an input image that influence those representations). The framework includes three complementary approaches: Sharing Ratio Decomposition for local explanations, Concept-Anchored Feature Explanation for connecting abstract features to visible pixels, and Interlayer Concept Graph for tracking how representations evolve through the model's layers. The authors demonstrate that their approach provides clearer, more reliable explanations of model decisions across different vision architectures (ResNet50, VGG16, and Vision Transformers) compared to existing methods.