Summary of Unlearning In- Vs. Out-of-distribution Data in Llms Under Gradient-based Method, by Teodora Baluta and Pascal Lamblin and Daniel Tarlow and Fabian Pedregosa and Gintare Karolina Dziugaite
Unlearning in- vs. out-of-distribution data in LLMs under gradient-based method
by Teodora Baluta, Pascal Lamblin, Daniel Tarlow, Fabian Pedregosa, Gintare Karolina Dziugaite
First submitted to arxiv on: 7 Nov 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 paper formalizes a metric to evaluate the quality of machine unlearning in large language models (LLMs), which is crucial for solving the problem of removing the influence of selected training examples from a learned model. The authors assess the trade-offs between unlearning quality and performance, demonstrating that unlearning out-of-distribution examples requires more steps but generally offers a better balance overall. In contrast, unlearning in-distribution examples shows a rapid decay in performance as unlearning progresses. Additionally, the study examines how example memorization and difficulty affect unlearning under a classical gradient ascent-based approach. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Machine learning researchers are working on a new problem called “unlearning.” This means taking away what a model learned from certain training examples. It’s important to figure out how well this works in large language models. The goal is to create a way to measure how good the unlearning is. The authors of this paper came up with a metric to do just that and tested it on different types of examples. They found that getting rid of out-of-distribution examples takes more steps, but it’s usually better overall. However, they also discovered that getting rid of in-distribution examples quickly makes the model worse. Finally, the study looked at how hard or easy it is to unlearn something from a model and what this means for the approach used. |
Keywords
* Artificial intelligence * Machine learning
