Summary of Proper Dataset Valuation by Pointwise Mutual Information, By Shuran Zheng et al.
Proper Dataset Valuation by Pointwise Mutual Information
by Shuran Zheng, Xuan Qi, Rui Ray Chen, Yongchan Kwon, James Zou
First submitted to arxiv on: 28 May 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Computer Science and Game Theory (cs.GT)
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 tackles the crucial problem of evaluating data curation techniques in artificial intelligence. Current methods assess a model’s performance on specific benchmarks, which can inadvertently promote practices that make data similar to the test data, violating Goodhart’s law. To address this, the authors propose an information-theoretic framework measuring dataset quality by its informativeness about true model parameters using the Blackwell ordering. They introduce a novel method for estimating mutual information between datasets and test data by training Bayesian models on embedded data and computing posteriors. Experiments demonstrate that their evaluation assigns lower scores to data curation strategies reducing dataset informativeness, whereas traditional methods may favor overfitting approaches. This work has implications for AI model development and highlights the importance of informative data. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary In simple terms, this research paper is about making sure that the data used to train artificial intelligence models is good quality. Currently, people evaluate how well a model performs on specific tests, but this can actually make the data worse by favoring certain types of data over others. The authors propose a new way to measure how informative a dataset is and compare it to other datasets. They show that their method is better at detecting when data curation techniques compromise the quality of the training data. |
Keywords
* Artificial intelligence * Overfitting
