Exploring the Feasibility of Formal Methods in Machine Learning and Artificial Intelligence
Year: 2024
Type: preprint
Abstract: This study investigates the application of formal methods in ensuring the accuracy and reliability of machine learning systems. Traditional formal approaches have proven effective in guaranteeing the correctness of both hardware and software systems. In this work, we specifically focus on the verification and validation of machine learning systems using state-of-the-art formal methods. To begin, we provide a concise overview of existing formal approaches, highlighting their significance and benefits. Subsequently, we delve into the specific formal methods that have been developed to validate the crucial stages of data preparation and training in machine learning systems. These methods aim to ensure that the input data is appropriately processed and that the training process produces accurate and reliable models. Moving forward, we explore formal methods employed for the comprehensive verification of machine learning systems. This involves considering both partial and exhaustive techniques to thoroughly evaluate system behavior and performance. Furthermore, we conduct a thorough examination of research works dedicated to verifying support vector machines and decision tree ensembles, as these are common machine learning algorithms with widespread applications. In conclusion, based on our findings, we propose several potential future directions for the formal verification of machine learning systems. These directions encompass areas such as the development of enhanced formal methods tailored to specific types of machine learning models, the integration of formal techniques into existing machine learning frameworks, and the exploration of techniques for handling the inherent uncertainty and complexity associated with machine learning systems. By pursuing these avenues, we aim to advance the field of formal verification and contribute to the trustworthy deployment of machine learning technologies.
Source: HAL
Author: Krichen, Moez
Language: English
Subfield: Computational Theory and Mathematics
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4390875062
