Semi-supervised Learning with Generative Adversarial Networks (GANs)

Modern deep learning classifiers require a large volume of labeled samples to be able to generalize well. GANs have shown a lot of potential in semi-supervised learning where the classifier can obtain good performance with very few labeled data (Salimans et. al., 2016).

Overview

Results

Table below shows cross-validation accuracy of semi-supervised learning GAN for 1000 epochs when 10% and 100% of MNIST data is labeled.

Figure below shows cross-validation accuracy for 1000 epochs when 10% of data is labeled. As can be seen here, training has not yet reached a plateau which indicates further training could provide higher accuracy.

Figures below show some generated samples at different epochs of training when 10% of data is labeled:

Reference:

Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., and Chen, X. (2016). Improved Techniques for Training GANs. In advances in Neural Information Processing Systems (NIPS), pages 2226-2234 (http://papers.nips.cc/paper/6125-improved-techniques-for-training-gans.pdf)