Evaluating the Effects of L1 and L2 Regularisation Placement in GAN Architectures

Kyaw Shwe, Zuu Zuu (2025) Evaluating the Effects of L1 and L2 Regularisation Placement in GAN Architectures. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (1MB) | Preview
Preview	PDF (Configuration Manual) Download (787kB) | Preview

Abstract

Generative Adversarial Networks (GANs) are perceptive to hyperparameter configurations and also unstable during training. Although regularisation is a typical approach to deep learning to help alleviate overfitting and enhance generalisation, its potential in GAN training has not been thoroughly investigated, particularly in regard to placement and type. This paper is a systematic exploration of the impact of L1 and L2 regularisation as positioned in various points of the GAN architecture such as the output of the generator, input of the discriminator and the weights of both networks. Experiments were first conducted on the DCGAN framework using the CIFAR-10 dataset, covering different configurations evaluated with Frechet Inception Distance (FID), Inception Score (IS), loss curve analysis, and visual inspection. The best performing configuration, applying L1 weight regularisation to the early layers of the discriminator (E5 L1 DiscW), achieved the lowest FID among all tested variants, though the improvement margin was modest. This configuration was successfully transferred to a conditional GAN, resulting in modest gains over the CGAN baseline. However, while some configurations improved FID and IS, visual inspection revealed only subtle perceptual differences in image quality. The results indicate that L1 and L2 regularisation may lead to small gains in stability, but it depends significantly on the architecture, the type of penalty and where it is inserted in the GAN. Future work should explore adaptive penalties, larger datasets, and modern GAN variants to realise more substantial gains in generative performance.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Singh, Jaswinder UNSPECIFIED
Subjects:	Q Science > QA Mathematics > Electronic computers. Computer science T Technology > T Technology (General) > Information Technology > Electronic computers. Computer science T Technology > T Technology (General) > Information Technology
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	01 Jul 2026 11:31
Last Modified:	01 Jul 2026 11:31
URI:	https://norma.ncirl.ie/id/eprint/9435

Actions (login required)

View Item