In plain words: A generator and a judge network compete, and new training tricks sharpen that contest to make more realistic images and learn better from few labels. People could not tell generated digits from real ones, and guessed wrong 21.3% of the time on color photos.
Abstract · Improved Techniques for Training GANs
We present a variety of new architectural features and training procedures that we apply to the generative adversarial networks (GANs) framework. We focus on two applications of GANs: semi-supervised learning, and the generation of images that humans find visually realistic. Unlike most work on generative models, our primary goal is not to train a model that assigns high likelihood to test data, nor do we require the model to be able to learn well without using any labels. Using our new techniques, we achieve state-of-the-art results in semi-supervised classification on MNIST, CIFAR-10 and SVHN. The generated images are of high quality as confirmed by a visual Turing test: our model generates MNIST samples that humans cannot distinguish from real data, and CIFAR-10 samples that yield a human error rate of 21.3%. We also present ImageNet samples with unprecedented resolution and show that our methods enable the model to learn recognizable features of ImageNet classes.
Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, Xi Chen
arXiv:1606.03498 · cs.LG, cs.CV, cs.NE · submitted Jun 10, 2016
abstract · pdf · html
One of the great hopes of the current deep learning boom is that somehow we will develop unsupervised or at least semi-supervised techniques which can perform close to the great results that are being seen with supervised learning.
Adversarial Networks is one of the more likely routes to semi-supervised. There is also a lot of interesting work in combining Bayesian optimization techniques with Deep Networks to develop one-shot learning[1][2]. Some of this was (very broadly) in response to the the one-shot learning paper coming out of (I've forgotten!!) where the authors are famously doubtful about the utility of Deep Learning, and showed somewhat competitive results on MNIST. (I can't remember who it was - there have been HN discussions about the group. Sorry!!)
Both OpenAI and DeepMind are following roughly similar paths here (no surprise really), and the results are looking really good.
[1] http://arxiv.org/abs/1603.05106
[2] http://arxiv.org/abs/1606.04080