Summary of Experts Don’t Cheat: Learning What You Don’t Know by Predicting Pairs, By Daniel D. Johnson et al.
Experts Don’t Cheat: Learning What You Don’t Know By Predicting Pairs
by Daniel D. Johnson, Daniel Tarlow, David Duvenaud, Chris J. Maddison
First submitted to arxiv on: 13 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 proposes a strategy to quantify the epistemic uncertainty of generative models by training them to predict pairs of independent responses drawn from the true conditional distribution. This approach allows the model to “cheat” by observing one response while predicting the other, and measuring how much it cheats provides an estimate of the remaining gaps between the model’s predictions and the true distribution. The authors prove that being good at cheating is equivalent to being second-order calibrated, enabling the construction of provably-correct frequentist confidence intervals for the true conditional distribution and detection of incorrect responses with high probability. This method is demonstrated to accurately estimate epistemic uncertainty across various tasks, including image classification, language modeling, and navigation. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper helps us understand how well a machine learning model knows what it’s doing by training it on pairs of similar things. We then let the model “cheat” by looking at one thing while trying to predict another. By measuring how much it cheats, we can figure out where the model is unsure or doesn’t know something. This helps us make better predictions and avoid mistakes. The authors show that this approach works well in different situations. |
Keywords
* Artificial intelligence * Image classification * Machine learning * Probability
