Verifying AI systems by extracting automata via learning

通过学习提取自动机来验证人工智能系统

• 批准号: 2791125
• 金额: --
• 依托单位: City, University of London
• 依托单位国家: 英国
• 项目类别: Studentship
• 财政年份: 2023
• 资助国家: 英国
• 起止时间: 2023 至 无数据
• 项目状态: 未结题
• 来源: https://gtr.ukri.org/projects?ref=studentship-2791125
• 关键词: Verifying; AI; systems; extracting; automata

Ensuring AI programs behave the way they are expected to, has become of utmost importance. However verifying AI systems is a challenge, both from a theoretical and from a practical point of view. The project aims at investigating how to extract mathematical models (such as automata) from AI systems (such as neural network) to capture their behaviour. The project will focus on developing learning algorithms for models of automata such as weighted automata and on applying them to extract models from recurrent neural networks - or other AI systems - capturing some of their behaviour to guarantee safety properties.

确保人工智能程序按照预期的方式运行，已经变得至关重要。然而，无论是从理论还是从实践的角度来看，验证人工智能系统都是一个挑战。该项目旨在研究如何从人工智能系统（如神经网络）中提取数学模型（如自动机）以捕获其行为。该项目将专注于开发自动机（如加权自动机）模型的学习算法，并将其应用于从循环神经网络或其他人工智能系统中提取模型，捕获它们的一些行为以保证安全属性。