Summary of Fairdedup: Detecting and Mitigating Vision-language Fairness Disparities in Semantic Dataset Deduplication, by Eric Slyman and Stefan Lee and Scott Cohen and Kushal Kafle
FairDeDup: Detecting and Mitigating Vision-Language Fairness Disparities in Semantic Dataset Deduplication
by Eric Slyman, Stefan Lee, Scott Cohen, Kushal Kafle
First submitted to arxiv on: 24 Apr 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
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 research paper explores the impact of dataset deduplication techniques on reducing the cost of training Vision-Language Pretrained (VLP) models without compromising performance. The study focuses on pruning image-caption datasets collected from the web, which often harbor social biases that can be perpetuated in trained models. To mitigate these negative effects, the authors introduce FairDeDup, a modified version of the SemDeDup algorithm, and evaluate its fairness metrics on the FairFace and FACET datasets. The results show that FairDeDup leads to improved fairness while maintaining zero-shot performance on CLIP benchmarks. |
Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper looks at how we can make AI models fairer by reducing biased data in training. Right now, many AI models are trained on big collections of images and words from the internet. These collections often contain biases against certain groups of people, like women or minorities. This means that these biases can get passed down to the AI model, making it unfair. The researchers developed a new way to remove some of this biased data before training the AI models. They tested this method on different AI models and found that it made them fairer without sacrificing their ability to perform well. |
Keywords
* Artificial intelligence * Pruning * Zero shot