Query Languages for Machine Learning Models

Emerging challenges in machine learning (ML), such as explainability and verification, underscore the growing need for declarative query languages that enable users to extract relevant information from ML models and adapt it to diverse application-specific requirements. These query languages offer several advantages: they provide flexibility in information extraction, establish clear syntax and semantics for queries, and pave the way for query optimization. In this talk, we survey two recent proposals for query languages tailored to ML models—one designed for discrete classification models and another for real-valued models. We demonstrate how these languages can express meaningful queries over ML models, and we analyze their expressiveness and evaluation complexity. Our goal is to foster a productive discussion on advancing the development of practical query languages for ML models that can be effectively applied across a wide range of scenarios.6