Comparative Analysis of Quantum and Classical Convolutional Autoencoders on the MNIST Dataset
Year: 2024
Type: dissertation
Abstract: The growing popularity of machine learning and the increasing need for performance and data has led to increased interest in finding new solutions and tools to achieve more with less. In this study, we explore the usage of quantum computing on this issue, specifically how random quantum circuits can be integrated into convolutional layers in convolutional autoencoders, creating so-called quanvolutional layers. We replicate experiments from previous works and expand upon them to further investigate how different properties of a quanvolutional layer relate to better performance. The study covers how different kernel sizes, different parameter dimensions, and trainable quantum parameters affect loss curves. Our results, which are in line with our hypothesis, show a continued lack of quantum benefit on the research of quantum convolutional neural networks. The results do however also point to some interesting characteristics of quanvolutional layers that might constitute promising leads in achieving quantum advantage, such as trainable quantum parameters.
Authors: Astorga Castillo, Dante, Friman, Lukas
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W6980671099