Is Neural Network Verification Useful and What Is Next?

Johnson, Taylor T.

Permalink

https://hdl.handle.net/2142/130315 Copy

Description

Title

Is Neural Network Verification Useful and What Is Next?

Author(s)

Johnson, Taylor T.

Issue Date

2025-09-17

Keyword(s)

• Neural network verification
• Safe AI
• Neural networks
• Machine learning
• Artificial intelligence
• Formal methods

Abstract

Neural network verification is an instantiation of the classical formal verification problem, which is, given a model of a system and a specification, both with precise syntax and semantics, prove that the model satisfies the specification. The neural network verification problem instantiates this generic problem, where the model is a neural network represented as a mathematical function and the specification is typically defined as subsets over the input and output spaces of the function, where the model satisfies the specification if all inputs map only to desired outputs. The past decade or so has witnessed significant progress on this problem, such as improved scalability (e.g., in the sizes of the neural network and specifications), application on a variety of machine learning tasks, and important case studies in critical domains ranging from autonomy to medicine, among other advancements. However, is this progress so far useful, what are the current strengths and weaknesses of results in this research area, and what’s next for this research area and related problems? This position presentation will take a critical view of this research area discussing these aspects, highlighting where it is and is not useful, and present suggestions for what problems to investigate next.

Publisher

Allerton Conference on Communication, Control, and Computing

Series/Report Name or Number

2025 61st Allerton Conference on Communication, Control, and Computing Proceedings

ISSN

2836-4503

Type of Resource

Text

Genre of Resource

Conference Paper/Presentation

Language

eng

Handle URL

https://hdl.handle.net/2142/130315&&

Copyright and License Information

Copyright 2025 is held by Taylor T. Johnson.

Owning Collections