Summary of Revisiting the Dataset Bias Problem From a Statistical Perspective, by Kien Do et al.
Revisiting the Dataset Bias Problem from a Statistical Perspective
by Kien Do, Dung Nguyen, Hung Le, Thao Le, Dang Nguyen, Haripriya Harikumar, Truyen Tran, Santu Rana, Svetha Venkatesh
First submitted to arxiv on: 5 Feb 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 paper investigates the “dataset bias” problem from a statistical standpoint, identifying the strong correlation between class attributes and non-class attributes as the primary cause. The study finds that this correlation is incorporated into model parameters when trained using maximum log-likelihood (MLL), leading to poor generalization on unbiased test data. To mitigate dataset bias, the authors propose two methods: weighting the objective by inverse probability or sampling with a weight proportional to it. The first method proves more stable and effective in practice, while the second is statistically equivalent but less computationally efficient. The paper also establishes a connection between debiasing and causal reasoning, providing theoretical foundations for the approach. Experiments on various biased datasets demonstrate the superiority of the proposed method over existing techniques. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper explores how dataset bias affects machine learning models. It shows that when we train models using certain methods, they can become “biased” towards specific characteristics in the training data. This means that the models won’t generalize well to new data that doesn’t have those same characteristics. The authors suggest two ways to fix this problem: either adjust the objective function or sample the training data differently. They also explain how their approach is related to understanding cause-and-effect relationships, which adds credibility to their method. By testing their ideas on different datasets, they show that their approach works better than other methods in most cases. |
Keywords
* Artificial intelligence * Generalization * Log likelihood * Machine learning * Objective function * Probability
