Understanding an AI system can take many forms. For instance, we might want to know when and how the system relies on sensitive or spurious features, identify systematic errors in its predictions, or learn how to modify the training data and model architecture to improve accuracy and robustness. Today, answering these types of questions often involves significant human effort—researchers must formalize their question, formulate hypotheses about a model’s decision-making process, design datasets on which to evaluate model behavior, then use these datasets to refine and validate hypotheses. As a result, this type of understanding is slow and expensive to obtain, even about the most widely used models.
Automated Interpretability approaches have begun to address the scalability problem. Recently, such approaches have used pretrained language models like GPT-4 (in Bills et al. 2023) or Claude (in Bricken et al. 2023) to generate explanations of features. In earlier work, we introduced MILAN (Hernandez et al. 2022), a captioner-like model trained on human feature annotations that takes as input a feature visualization and outputs a description of the feature’s functionality based on the visualization. But automated approaches that use learned models to label features leave something to be desired: they are primarily tools for hypothesis generation (Huang et al. 2023), they characterize behavior on a limited set of inputs, and they are often low precision.
Our current line of research aims to build tools that help users understand models, while combining the flexibility of human experimentation with the scalability of automated techniques. We take an approach based on automating scientific experimentation on models -- describe processes underlying data they generate. In Schwettmann et al. 2023, we introduced the interactive Automated Interpretability Agent paradigm, where LM-based agent interactively probe systems to explain their behavior.Vision-language backbone and a sophisitcated experiments on other systems (see many more examples in our neuron viewer).
@@ -157,7 +157,7 @@