Abstraction-based neural network analysis by probabilistic abstract interpretation
File(s)
Author(s)
Zhang, Zhuofan
Type
Thesis or dissertation
Abstract
Probabilistic abstract interpretation is a theory that is used to extract particular properties of a computer program when it is not feasible to test every input. In this thesis, we apply the theory on neural networks for the same purpose: to analyze density distribution flow of all possible inputs of a neural network when a network has uncountably many or countable but infinitely many inputs. We show how this theoretical framework works on different types of neural networks. We then discuss different abstract domains as part of the analysis and their corresponding Moore-Penrose pseudo-inverses together with abstract transformers. We also give experimental illustrations and evaluations.
Version
Open Access
Date Issued
2025-11-24
Date Awarded
2026-08-01
Copyright Statement
Attribution-NonCommercial 4.0 International Licence (CC BY-NC)
License URL
Advisor
Wiklicky, Herbert
Publisher Department
Department of Computing
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)