Answers are grounded in this paper and cite their sources.
We propose a novel GAN training scheme that can handle any level of labeling a unified manner. Our scheme introduces a form of artificial labeling that incorporate manually defined labels, when available, and induce an between them. To define the artificial labels, we exploit the that neural network generators can be trained more easily to map latent vectors to data with semantic similarities, than across separate. We use generated data samples and their corresponding artificial labels to train a classifier. The classifier is then used to-label real data. To boost the accuracy of the self-labeling, we also use exponential moving average of the classifier. However, because the might still make mistakes, especially at the beginning of the, we also refine the labels through self-attention, by using the of real data samples only when the classifier outputs a high probability score. We evaluate our approach on CIFAR-10, STL-10 SVHN, and show that both self-labeling and self-attention consistently the quality of generated data. More surprisingly, we find that the scheme can even outperform class-conditional GANs.
No takes yet. Share an insight, caveat, or question.
Watanabe et al. (2021) studied this question.