Figureure 2 · Overview of our ViSAE toolbox for interpreting ViT inner workingsFigure 2. Overview of our ViSAE toolbox for interpreting ViT inner workings. Left: Motivated by the human visual cortex hierarchy, we construct a probing suite (64K images + 16K concepts) for SAE training and interpretation. Middle: Our top-down concept reading and bottom-up concept circuit tracing algorithms. Right: Our mechanistic view of ViT inner workings enables various downstream applications, such as concept localization, failure mode analysis, and model steering.这张图概括 Inside the Visual Mind 的整体方法流程。阅读时先看模块之间传递的训练信号,再看作者如何把目标拆成可优化的子问题。Figureure 1 · Existing interpretable machine learning methods (IML) mainly identify Figure 1. Existing interpretable machine learning methods (IML) mainly identify where the evidence is, while our concept circuits reveal how concepts interact across layers to support a prediction.这张图/表用于判断 Inside the Visual Mind 的实验收益来自哪里。重点看替换、消融或跨模型设置下趋势是否一致,而不是只看单个最高分。