Summary of Taming the Tail in Class-conditional Gans: Knowledge Sharing Via Unconditional Training at Lower Resolutions, by Saeed Khorram et al.
Taming the Tail in Class-Conditional GANs: Knowledge Sharing via Unconditional Training at Lower Resolutions
by Saeed Khorram, Mingqi Jiang, Mohamad Shahbazi, Mohamad H. Danesh, Li Fuxin
First submitted to arxiv on: 26 Feb 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The proposed study aims to improve the training of generative adversarial networks (GANs) with long-tailed training data. GANs tend to favor classes with more samples, leading to low-quality and less diverse samples in tail classes. The researchers propose a method for knowledge sharing, allowing tail classes to borrow from classes with more abundant training data. They modify existing class-conditional GAN architectures to train lower-resolution layers unconditionally and reserve class-conditional generation for higher-resolution layers. Experimental results show significant improvements in generated image diversity and fidelity on several long-tail benchmarks. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Long-tailed training data is a common problem in generative adversarial networks (GANs). The problem is that GANs tend to focus on classes with more samples, leaving the less represented classes looking bad. Researchers are working on ways to help GANs generate better images from these long-tailed datasets. One way to do this is by sharing information between classes. This study proposes a simple but effective method for doing just that. They modify existing GAN architectures to train lower-resolution layers without considering the class, and then use class information when generating higher-resolution details. The results show that their method can generate more diverse and realistic images than current methods. |
Keywords
* Artificial intelligence * Gan
