{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/improving-gan-with-neighbors-embedding-and","title":"Improving GAN with neighbors embedding and gradient matching","arxiv_id":"1811.01333","date":"2018-11-04","proceeding":null,"authors":["Ngoc-Trung Tran","Tuan-Anh Bui","Ngai-Man Cheung"],"abstract":"We propose two new techniques for training Generative Adversarial Networks\n(GANs). Our objectives are to alleviate mode collapse in GAN and improve the\nquality of the generated samples. First, we propose neighbor embedding, a\nmanifold learning-based regularization to explicitly retain local structures of\nlatent samples in the generated samples. This prevents generator from producing\nnearly identical data samples from different latent samples, and reduces mode\ncollapse. We propose an inverse t-SNE regularizer to achieve this. Second, we\npropose a new technique, gradient matching, to align the distributions of the\ngenerated samples and the real samples. As it is challenging to work with\nhigh-dimensional sample distributions, we propose to align these distributions\nthrough the scalar discriminator scores. We constrain the difference between\nthe discriminator scores of the real samples and generated ones. We further\nconstrain the difference between the gradients of these discriminator scores.\nWe derive these constraints from Taylor approximations of the discriminator\nfunction. We perform experiments to demonstrate that our proposed techniques\nare computationally simple and easy to be incorporated in existing systems.\nWhen Gradient matching and Neighbour embedding are applied together, our GN-GAN\nachieves outstanding results on 1D/2D synthetic, CIFAR-10 and STL-10 datasets,\ne.g. FID score of $30.80$ for the STL-10 dataset. Our code is available at:\nhttps://github.com/tntrung/gan","url_abs":"http://arxiv.org/abs/1811.01333v1","url_pdf":"http://arxiv.org/pdf/1811.01333v1.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"improving-gan-with-neighbors-embedding-and","repo_url":"https://github.com/tntrung/gan","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"tf","reach":{"status":"ok","spdx":"GPL-3.0"}}],"tasks":[],"methods":[{"method_slug":"convolution","method_name":"Convolution"}],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1811.01333","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}