Multimodal Interpretability in 2024

In 2024, the field of multimodal interpretability is evolving, with a focus on mechanistic and causal interpretability rather than traditional methods like saliency maps. The author, who has experience in AI safety and video understanding, discusses various approaches to understanding how multimodal models, particularly Vision-Language Models (VLMs), operate. Key methods include circuit-based approaches that analyze the computational subgraph of models, and techniques that leverage the shared text-image space to interpret vision models using text embeddings. The TEXTSPAN algorithm is highlighted for its ability to create text-labeled bases for vision encoder outputs, enhancing interpretability. The author also explores the challenges of applying text-based interpretation methods to models that lack a shared text-image space, proposing solutions such as training aligned text embeddings or using adapters to map embeddings between different models. The discussion emphasizes the potential for these methods to improve model customization and auditing, while also acknowledging the need for further research to validate their effectiveness and address critiques regarding the representational adequacy of neurons and features.

- Multimodal interpretability is shifting towards mechanistic and causal methods.

- Circuit-based approaches and shared text-image space techniques are key focus areas.

- The TEXTSPAN algorithm enhances interpretability by linking vision model outputs to text descriptions.

- Challenges exist in applying text-based methods to models without aligned text embeddings.

- Proposed solutions include training aligned text embeddings and using lightweight adapters for mapping.