Reproduce “Semi-Supervised Learning with GANs: Revisiting Manifold Regularization” on OpenResearch

About this paper

Bruno Lecouat, Chuan-Sheng Foo, Houssam Zenati, Vijay R. Chandrasekhar

The authors add manifold regularization to a feature-matching GAN by sampling nearby images through its generator and penalizing changes in classifier predictions across latent-space perturbations. On CIFAR-10, with 4,000 labeled examples and no augmentation, test error falls from 15.5 ± 0.35% to 14.45 ± 0.21%; BadGAN reports a slightly lower 14.41 ± 0.30%, so the result supports improvement over the authors’ own unregularized implementation, not best performance among GAN methods.

Key insights

• GANs' inherent ability to model the data manifold can be directly leveraged to simplify and efficiently implement manifold regularization for semi-supervised learning.
• Approximating the Laplacian norm via Monte Carlo sampling and stochastic finite differences in the latent space provides a computationally tractable and scalable approach to manifold regularization.
• The regularization encourages the discriminator's output to be invariant to small perturbations in the generator's latent space, promoting local smoothness on the learned data manifold.

Summary by alphaXiv.