Summary of Debiasing Large Vision-language Models by Ablating Protected Attribute Representations, By Neale Ratzlaff et al.
Debiasing Large Vision-Language Models by Ablating Protected Attribute Representations
by Neale Ratzlaff, Matthew Lyle Olson, Musashi Hinck, Shao-Yen Tseng, Vasudev Lal, Phillip Howard
First submitted to arxiv on: 17 Oct 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Computation and Language (cs.CL); 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 paper proposes a novel framework for Large Vision Language Models (LVLMs) to mitigate societal biases in their responses. The current models, such as LLaVA, are trained on datasets containing biases and respond differently when presented with images of people from various demographics. To address this issue, the proposed debiasing method directly ablates biased attributes during text generation, preventing the model from generating text related to protected attributes or representing them internally. This approach requires minimal training (1000 samples) and maintains captioning performance on real data like COCO. The results show that debiasing can be achieved without sacrificing model performance, as evidenced by similar accuracy between a debiased LVLM and its biased counterpart. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper aims to make Large Vision Language Models fairer by reducing their responses to specific demographics. Current models are trained on datasets containing biases, which affects how they respond to different people’s images. The authors suggest a new way to fix this issue without retraining the model. They “remove” biased words during text generation to avoid mentioning certain groups or attributes. This method works with just 1000 examples and keeps the model’s performance good on real data like COCO. |
Keywords
* Artificial intelligence * Text generation
