10/18/2024
For the fourth consecutive year, ECE Professor Huan Zhang has led his team to victory at the 5th International Verification of Neural Networks Competition (VNN-COMP) with an open-source software toolkit that aims to solve the challenges of neural network verification quickly and efficiently.
Written by
Researchers at The Grainger College of Engineering are pioneering new algorithms to verify the trustworthiness of AI-based systems. For the fourth consecutive year, ECE Professor Huan Zhang has led his team to victory at the 5th International Verification of Neural Networks Competition (VNN-COMP) with an open-source software toolkit that aims to solve the challenges of neural network verification quickly and efficiently.
Neural networks are widely used in artificial intelligence systems, powering many AI applications such as autonomous driving and chatbots. The emerging field of neural network verification aims to mathematically prove that a neural network will behave as intended in all possible scenarios, providing a safeguard to ensure trustworthy AI systems as technology advances.
“Just as we rigorously test and verify critical software systems, neural network verification provides a way to formally guarantee that an AI system will produce the correct outputs and avoid undesirable behavior,” Zhang explains. “This is especially important for AI systems being used in high-stakes domains like self-driving vehicles, medical diagnosis, and financial trading, where unexpected or incorrect outputs could lead to serious real-world consequences.”
Zhang’s team used their open-source software toolkit, the alpha-beta-CROWN verifier, to outperform seven other teams and win the competition. The team began developing this tool in 2020, and this is the fourth consecutive year that they have participated in and won VNN-COMP. Other teams during the past four years come from institutions including Stanford University, Carnegie Mellon University, Stony Brook University, Vanderbilt University, UC Irvine, Oxford and ZTH Zurich.
In the competition, each team’s verification tools are run on a set of standardized benchmarks containing verification problems (mathematical challenges) for the tool to solve. The team that solves the most challenges wins the competition. The 2024 competition marks the first time that Zhang’s team won first place among all 21 benchmarks across both competition tracks, outperforming all other participating teams in every benchmark.
As AI-based systems advance and become increasingly embedded in our society, it is vital to ensure that neural network behaviors are well understood, reliable and predictable. Beyond the competition, Zhang’s team is actively working on applying neural network verification to enable formal guarantees for AI in different engineering domains, and they have recently published a paper on verifying neural networks in control at the International Conference on Machine Learning (ICML 2024). Zhang’s team will continue to advance towards the goal of guaranteed safe AI in mission-critical domains.
Grainger Engineering Affiliations
Huan Zhang is an Illinois Grainger Engineering professor of Electrical and Computer Engineering and is affiliated with the Coordinated Science Laboratory and the Siebel School of Computing and Data Science.