AI for V&V
Chair: Peter Popov
Program Verification for Rigorous Analysis of Decision Tree Ensembles against Specifications
Iat Tou Leong, Aleksandar Avdalovic, Raul Barbosa
As machine learning continues to integrate into critical
systems, ensuring the reliability and correctness of these
models becomes essential. Decision tree ensembles, which
combine multiple decision trees to improve performance and
robustness, present unique challenges for verification due
to their complexity. This paper explores the application of
theorem proving techniques for the formal verification of
decision tree ensembles against specified requirements.
Theorem proving can identify and mitigate potential
specification violations, enhancing the trustworthiness and
safety of machine learning ensembles. Two case studies and
experimental results show the effectiveness of the
approach, highlighting its potential to serve as a critical
tool in the deployment of reliable machine learning
systems. Moreover, the paper describes an evaluation of the
complexity of the verification process, focusing on the
computational considerations and feasibility of applying
theorem proving to random forests.
Overview
Program