Summary of Training Unbiased Diffusion Models From Biased Dataset, by Yeongmin Kim et al.
Training Unbiased Diffusion Models From Biased Dataset
by Yeongmin Kim, Byeonghu Na, Minsang Park, JoonHo Jang, Dongjun Kim, Wanmo Kang, Il-Chul Moon
First submitted to arxiv on: 2 Mar 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Computer Vision and Pattern Recognition (cs.CV)
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 This paper addresses the crucial issue of dataset bias in diffusion models, which directly affects generated output quality and proportion. The authors propose time-dependent importance reweighting to mitigate latent bias, demonstrating that this approach is more precise than previous methods. The proposed method enables a tractable form of the objective function for regenerating unbiased data density. Additionally, the paper establishes a connection between score-matching and traditional score-matching, providing theoretical support for convergence to an unbiased distribution. Experimental results on CIFAR-10, CIFAR-100, FFHQ, and CelebA datasets with various bias settings show that the proposed method outperforms baselines. The authors’ code is available online. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This research paper is about making sure computer-generated images don’t have hidden biases. When we use biased data to create new pictures or videos, it can affect how realistic they are and what kind of messages they send. The scientists propose a way to fix this problem by adjusting the importance of certain parts of the data. They show that their method is more accurate than others and works well on different types of images. This helps ensure that computer-generated content is fair and unbiased. |
Keywords
* Artificial intelligence * Diffusion * Objective function
