Summary of Discovering Bias in Latent Space: An Unsupervised Debiasing Approach, by Dyah Adila et al.
Discovering Bias in Latent Space: An Unsupervised Debiasing Approach
by Dyah Adila, Shuai Zhang, Boran Han, Yuyang Wang
First submitted to arxiv on: 5 Jun 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: None
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 SteerFair method tackles the issue of foundation models being highly sensitive to prompt variations, which is often due to their preference or bias towards specific input characteristics. This bias can lead to superficial, non-meaning-altering changes affecting model performance. To rectify this, SteerFair finds the bias direction in the model’s internal representation and steers activation values away from it during inference. The approach exploits the observation that bias often adheres to simple association rules and constructs demonstrations of these rules from unlabeled samples. SteerFair significantly reduces instruction-tuned model performance variance across prompt modifications on three benchmark tasks, outperforming a supervised baseline with 100 labels by an average of 10.86% accuracy points and 12.95 score points. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Foundation models are great at answering questions, but they can be very sensitive to tiny changes in the question itself. This makes it hard for them to give good answers when the question is changed just a little bit. One reason this happens is that the model has a bias or preference for certain types of information. For example, if the model looks at pictures and sees an animal first, it might think that’s what the question is asking about even if the rest of the picture doesn’t have anything to do with animals. The researchers came up with a new way to fix this problem called SteerFair. It works by looking at how the model is biased and then “steering” its answers away from those biases. This helps the model give more accurate answers, even when the question is changed just a little bit. |
Keywords
» Artificial intelligence » Inference » Prompt » Supervised
