Automated Interpretation of Machine Learning Models

Abstract

As machine learning (ML) models are increasingly deployed in production, there’s a pressing need to ensure their reliability through auditing, debugging, and testing. Interpretability, the subfield that studies how ML models make decisions, aspires to meet this need but traditionally relies on human-led experimentation or is based on human priors about what the model has learned. In this thesis, I propose that interpretability should evolve alongside ML by adopting automated techniques that use ML models to interpret ML models. This shift towards automation allows for more comprehensive analyses of ML models without requiring human scrutiny at every step, and the effectiveness of these methods should improve as the ML models themselves become more sophisticated. I present three examples of automated interpretability approaches: using a captioning model to label features of other models, manipulating a ML model’s internal representations to predict and correct errors, and identifying simple internal circuits through approximating the ML model itself. These examples lay the groundwork for future efforts in automating ML model interpretation.