Improving inverse abstraction based neural network verification using automated machine learning techniques

Matthias könig

PhD at Leiden University

Abstract: This project seeks to advance the state of the art in formal neural network verification. Formal neural network verification methods check whether a trained neural network, for example an image classifier, satisfies certain properties or guarantees regarding its behaviour, such as correctness, robustness, or safety, under various inputs and conditions. The main goal of this project is to study incomplete verification methods, a type of verification methods that are generally fast but may not always find a solution. Specifically, we are going to investigate if we can improve their performance using automated algorithm configuration techniques.

Keywords: neural network verification, automated algorithm configuration, portfolio construction, incomplete verification methods, scalability

Scientific area: Artificial Intelligence

Matthias is a fourth-year PhD candidate with an interest in AI Safety and automated machine learning. More specifically, his research focuses on combining methods for automated algorithm configuration with formal verification of neural networks.