Summary of How Data Inter-connectivity Shapes Llms Unlearning: a Structural Unlearning Perspective, by Xinchi Qiu et al.
How Data Inter-connectivity Shapes LLMs Unlearning: A Structural Unlearning Perspective
by Xinchi Qiu, William F. Shen, Yihong Chen, Meghdad Kurmanji, Nicola Cancedda, Pontus Stenetorp, Nicholas D. Lane
First submitted to arxiv on: 24 Jun 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
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 A novel method for compiling structural datasets, PISTOL, is proposed to address the limitations of existing approaches and benchmarks in large language models (LLMs) unlearning. By leveraging the structured nature of contractual relationships, PISTOL enables insights into the impact of data inter-connectivity on unlearning effectiveness, provides precise ground truths for evaluation, and mitigates confounding risks by not requiring input from pre-trained LLMs. The authors demonstrate that data inter-connectivity increases unlearning difficulty in both pre-trained and fine-tuned models, with a positive correlation between knowledge graph density and unlearning difficulty. Moreover, balancing retaining performance across all domains becomes challenging when the to-be-forgotten data is skewed towards one domain. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary LLMs can forget information they’ve learned, but existing methods assume this information is independent, like individual words. This paper introduces a new way to organize data that reflects real-world relationships. It shows how these connections affect how well LLMs can forget and how hard it is for them to do so. The results suggest that when the connections between pieces of information are strong, it’s harder for the models to forget. |
Keywords
* Artificial intelligence * Knowledge graph
