Carnegie Mellon University researchers led a team to victory in the 2021 International Verification of Neural Networks Competition with an open-source tool that can provide a guarantee of the behavior of a critical part of modern artificial intelligence.

Deep neural networks are a critical part of modern artificial intelligence, powering applications from computer vision to natural language processing. But they are complex and unpredictable, posing challenges for areas where safety is paramount, like autonomous driving, aircraft autopilot and medical systems.

"Neural networks are often black-boxes, and it is hard to guarantee that they behave safely and predictably under noisy or malicious inputs," said Huan Zhang, a postdoctoral researcher in the Computer Science Department. "Neural network verification aims to provide provable guarantees for desired properties of neural networks, such as safety and robustness."

Zhang led the team, which included Zico Kolter, an associate professor in CMU’s School of Computer Science, and researchers from Northeastern University, Columbia University and UCLA. The team won with an open-sourced tool called α,β-CROWN (alpha-beta-CROWN). The collection of algorithms developed by the team are efficient and fast. They run in parallel to make full use of modern computing power. The tool is general and scalable and performed well on a variety of distinct tasks with diverse neural network structures

"α,β-CROWN works at speeds two to three orders of magnitude faster than tools developed just three years ago," Zhang said. "This has led to significant progress in neural network verification and could lower barriers to adoption in safety-critical fields."

The competition compared the scalability and speed of tools and how well they performed in eight scoring benchmarks, including image classification, control and database.

The team won this year's competition with a score of 779.2 out of 800, nearly 80 points higher than the second-place team from Imperial College London and nearly 200 points higher than the third-place teams, one from the University of Oxford and one from ETH Zurich and University of Illinois Urbana-Champaign.

This was the second year for the International Verification of Neural Networks Competition. Twelve teams competed this year, and while first place took home $400, second $300 and third $200, it wasn't about the money. Neural network verification is crucial to the further development of AI.

"The next major challenge in modern AI is to build more reliable and predictable systems that can be trusted in mission-critical environments," Zhang said. "Neural network verification is a significant step towards this goal because it provides a way to guarantee the behavior of complex deep neural networks."