Seeing Through Circuits: Faithful Mechanistic Interpretability for Vision Transformers
For interpretability researchers, this work extends mechanistic interpretability from language to vision models, offering a new tool to understand information routing in vision transformers.
The paper introduces Vi-CD, a method for discovering edge-based mechanistic circuits in vision transformers, enabling identification of class-specific circuits, circuits underlying typographic attacks in CLIP, and circuits for steering model behavior. Results show that these circuits provide actionable transparency into internal computations.
Transparency of neural networks' internal reasoning is at the heart of interpretability research, adding to trust, safety, and understanding of these models. The field of mechanistic interpretability has recently focused on studying task-specific computational graphs, defined by connections (edges) between model components. Such edge-based circuits have been defined in the context of large language models, yet vision-based approaches so far only consider neuron-based circuits. These tell which information is encoded, but not how it is routed through the complex wiring of a neural network. In this work, we investigate whether useful mechanistic circuits can be identified through computational graphs in vision transformers. We propose an effective method for Automatic Visual Circuit Discovery (Vi-CD) that recovers class-specific circuits for classification, identifies circuits underlying typographic attacks in CLIP, and discovers circuits that lend themselves for steering to correct harmful model behavior. Overall, we find that insightful and actionable edge-based circuits can be recovered from vision transformers, adding transparency to the internal computations of these models.