Year: 2025
Type: dissertation
Abstract: Diffusion models have emerged as powerful generative modeling frameworks, excelling at producing state-of-the-art high-quality, diverse data samples across various domains. Although slightly suffering from slow sampling speed compared to other predominant approaches, various methods such as implicit and latent methods have addressed this issue. In many machine learning algorithms, including diffusion models, the availability and size of training datasets is a critical factor in their performance. This presents challenges in limited-data scenarios commonly encountered in specialized fields such as with medical data. This thesis comprehensively surveys diffusion models, analyzing their mathematical foundations, architectural innovations, and application domains, while specifically addressing their capabilities and limitations in low-data contexts. Through empirical evaluations on the standard image datasets CIFAR-10 and MNIST, and medical tabular data, we systematically investigate how dataset size influences diffusion model performance. Furthermore, we explore their potential in augmenting scarce datasets through data synthesis, evaluating the effectiveness of diffusion-augmented classification tasks. Our findings indicate that in several limited-data scenarios, integrating synthetic samples generated by diffusion models enhances classification accuracy, underscoring their practical utility in real-world, data-constrained environments.
Authors: Lundquist, Oscar, Larsson, Kim
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7079202263