Reproduce “LatentAugment: Data Augmentation via Guided Manipulation of GAN's Latent Space” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Lorenzo Tronchin, Minh H. Vu, Paolo Soda, Tommy Löfstedt
LatentAugment introduces a data augmentation framework that uses guided manipulation of a pre-trained StyleGAN2's latent space to synthesize diverse and high-fidelity images. This approach demonstrably improves the generalization performance of deep learning models, achieving a 13.8% decrease in MAE for MRI-to-CT image translation compared to a baseline without augmentation.
Key insights
- Guided manipulation within a GAN's latent space can effectively address the "generative learning trilemma" by simultaneously enhancing diversity and fidelity for data augmentation.
- Semantic diversity, particularly through the latent loss term, is more impactful for improving downstream model generalization than simply enforcing fidelity, especially with a strong generator like StyleGAN2.
- The method's hyper-parameter optimization can be effectively performed in both task-dependent (e.g., MAE) and task-agnostic (e.g., F1 score for generative quality/diversity) manners.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.