Fast, formal, local eXplainable artificial intelligence for neural network

• Main Author: Wong, Laurentius Reynald
• Other Authors: Thomas Peyrin
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Engineering
• Online Access: https://hdl.handle.net/10356/176177
• Institution: Nanyang Technological University

The impact of artificial intelligence has transformed our daily lives, enabling us to complete tasks with ease. However, its application in critical areas such as autonomous vehicles, healthcare, and finance presents safety concerns. The lack of transparency in AI, functioning as a black box, leaves users with uncertainties and questions. This final year project builds upon the framework of Truth Table Net (TTNet), a novel rule-based model trained by a deep Convolutional Neural Network. TTNet currently provides global, formal interpretations of a model, which offer valuable insights into the overall model behavior. However, in many scenarios, local interpretations are preferred as they provide detailed explanations. The focus of this final year project is to develop a framework, V, that automates the computation of local interpretations for TTNet. Specifically, the project aims to generate contrastive explanations (CXP) and abductive explanations (AXP). V is then tested on several tabular datasets, including adult, breast cancer, mushroom, spambase, and diabetes datasets. In the experiments, V demonstrated promising results, exhibiting very low inference times across all datasets. This suggests that the methodologies used to create V can be implemented in various industries requiring formal verification. Additionally, limitations and potential future work are discussed.