Reproduce “Manifold regularization with GANs for semi-supervised learning” on OpenResearch

About this paper

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

The method uses a GAN to guide semi-supervised image classification, encouraging a classifier to keep its outputs stable along directions estimated from changes to the GAN’s latent input. On CIFAR-10 with 1,000 labeled examples, test error was 17.50% for the feature-matching GAN, 16.81% with ambient regularization, and 16.37% with manifold regularization, averaged over four runs without data augmentation. This result applies to that setting.

Key insights

• Efficient approximation of manifold regularization, focusing on the direction of manifold gradients via GAN-generated samples, effectively improves classifier performance in semi-supervised learning.
• Anisotropic manifold regularization, which considers data-dependent perturbations along specific manifold directions, yields superior performance compared to isotropic ambient regularization.
• There is a positive correlation between the quality of the GAN's generator (its ability to model the data manifold) and the effectiveness of the manifold regularization strategy on classification performance.

Summary by alphaXiv.