Summary of Images Speak Louder Than Words: Understanding and Mitigating Bias in Vision-language Model From a Causal Mediation Perspective, by Zhaotian Weng et al.
Images Speak Louder than Words: Understanding and Mitigating Bias in Vision-Language Model from a Causal Mediation Perspective
by Zhaotian Weng, Zijun Gao, Jerone Andrews, Jieyu Zhao
First submitted to arxiv on: 3 Jul 2024
Categories
- Main: Artificial Intelligence (cs.AI)
- Secondary: Computation and Language (cs.CL); 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 proposes a framework that uses causal mediation analysis to understand how biases are generated and propagated within vision-language models (VLMs). Specifically, it investigates how pre-trained VLMs learn biases by correlating gender information with specific objects or scenarios. The authors find that image features are the primary contributors to bias, accounting for 32.57% of the bias in the MSCOCO dataset and 12.63% in the PASCAL-SENTENCE dataset. They also identify that blurring gender representations within the image encoder reduces bias efficiently while minimizing performance loss or increased computational demands. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The researchers show that biases in VLMs are mainly caused by image features, which is different from previous methods that focused on modifying inputs and monitoring changes in output probability scores. The proposed framework allows for a comprehensive understanding of bias generation and propagation within VLMs, enabling more effective strategies to reduce bias. By identifying the direct effects of interventions on model bias and indirect effects mediated through different components, this study provides valuable insights into how biases are learned and perpetuated in pre-trained VLMs. |
Keywords
* Artificial intelligence * Encoder * Probability
