Automated Interpretability

Automated Interpretability Techniques

Interpretability Meta Metrics

Automated interpretability is a difficult task for models that we have only begun working on, it has already been very useful for quickly understanding dictionary learning features in a scalable fashion.

Towards Monosemanticity: Decomposing Language Models With Dictionary Learning

Mechanistic interpretability seeks to understand neural networks by breaking them into components that are more easily understood than the whole. By understanding the function of each component, and how they interact, we hope to be able to reason about the behavior of the entire network. The first step in that program is to identify the correct components to analyze.

https://transformer-circuits.pub/2023/monosemantic-features#appendix-automated

Rubric for human annotator (evaluating interpretability description)

Towards Monosemanticity: Decomposing Language Models With Dictionary Learning

https://transformer-circuits.pub/2023/monosemantic-features#appendix-rubric

Towards Monosemanticity: Decomposing Language Models With Dictionary Learning

https://transformer-circuits.pub/2023/monosemantic-features#global-analysis-interp

Natural language explanations based on individual neurons are not the best approach for LLM explainability. This claim begins with questions about whether natural language can faithfully describe how individual neurons store and process information. However, interventions on individual neurons showed no causal link between explanations and changes in model output.

arxiv.org

https://arxiv.org/pdf/2309.10312

Automated Interpretability

Rubric for human annotator (evaluating interpretability description)

Backlinks

Recommendations