Title: “Formal XAI”: Can we formally explain ML models?

Speaker: Shahaf Bassan    (bio) (bio)

Shahaf Bassan is a senior PhD student in the Katz Lab at the Hebrew University of Jerusalem specializing in explainable AI. His research focuses on developing explanation techniques with formally provable guarantees, at the intersection of explainability, formal verification, and ML theory. His work spans both theoretical foundations and practical applications. Sahaf has presented his research at leading conferences in formal verification (e.g., TACAS) and machine learning (e.g., ICML, ICLR). His research goal is to enhance trust in ML models by providing trustworthy, verifiable explanations.

When: Tuesday, 10 June 2025 at 1900 hrs (IST)  

Meeting Details: Zoom link, ID: 891 6409 4870, Passcode: 082194

Abstract:
The goal of explainability is to make sense of the decisions made by black-box ML models. Unfortunately, many existing explanation methods are heuristic, which makes them unreliable. In this talk, I will present our work on developing techniques that provide explanations with formal guarantees, ensuring their trustworthiness. These techniques often rely on formal verification, particularly neural network verification tools. In addition, we examine these explanations from a theoretical perspective - studying the computational challenges they pose and exploring ways to build practical tools that address these challenges and enable the generation of reliable explanations for ML models.