Semi-supervised generative adversarial networks (GANs) consist of neural networks tasked with learning the underlying distributions of each class of a target dataset, usually images. We approach the GAN learning problem from a novel perspective, one that is motivated by the famous Persian poet Rumi, who said, “The art of knowing is knowing what to ignore,” we provide the GAN with both positive data that it must learn to model, and negative samples that it must learn to avoid. The proposed formulation is able to generate realistic high-resolution images on datasets with positive to negative class ratios as low as 1:50.
References:
Asokan, S. and Seelamantula, C. S. “Teaching a GAN what not to Learn.” In Advances in Neural Information Processing Systems (NeurIPS), 2020
https://tinyurl.com/NeurIPSRumiGAN
https://github.com/DarthSid95/RumiGANs
https://papers.nips.cc/paper/2020/file/29405e2a4c22866a205f557559c7fa4b-Paper.pdf
