Reproduce “Data-Efficient GAN Training Beyond (Just) Augmentations: A Lottery Ticket Perspective” 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

Tianlong Chen, Yu Cheng, Zhe Gan, Jingjing Liu, Zhangyang Wang

A data-efficient framework for Generative Adversarial Networks is proposed, integrating the Lottery Ticket Hypothesis with data-level and a novel feature-level adversarial augmentation (AdvAug). This approach achieves state-of-the-art results in data-limited GAN training and enables competitive few-shot image generation without relying on large-scale pre-training.

Key insights

  • The Lottery Ticket Hypothesis extends to GANs in data-scarce regimes, where identified sparse "GAN tickets" are inherently more data-efficient and less prone to overfitting than dense models.
  • Adversarial augmentation applied to intermediate feature representations (AdvAug) acts as an effective regularization technique, consistently improving GAN performance by preventing overfitting.
  • Network sparsity, data-level augmentations, and the proposed feature-level AdvAug offer orthogonal and synergistic benefits, cumulatively enhancing data efficiency in GAN training.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.