Reproduce “Good Semi-supervised Learning that Requires a Bad GAN” 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

Zihang Dai, Zhilin Yang, Fan Yang, William W. Cohen, Ruslan Salakhutdinov

This research explains why generative adversarial networks are effective in semi-supervised learning by showing that a "complement generator" that produces diverse, low-density samples, rather than perfectly realistic ones, is optimal for improving discriminator classification boundaries. Their new framework achieved state-of-the-art semi-supervised classification error rates on MNIST, SVHN, and CIFAR-10, notably reducing the error rate on SVHN from 8.11% to 4.25%.

Key insights

  • A 'perfect' generator that exactly matches the true data distribution does not provide an inherent benefit for semi-supervised learning within the (K+1)-class discriminator framework.
  • Optimal semi-supervised learning requires a 'complement generator' that strategically produces diverse samples in low-density regions or the spaces between classes, as this forces the discriminator to learn robust, low-density decision boundaries.
  • The 'bad GAN' paradox is resolved: generators producing non-realistic, diverse, and low-density samples are not a weakness but a necessary condition for strong semi-supervised classification performance.

Summary by alphaXiv.

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