Summary of Towards Robust and Parameter-efficient Knowledge Unlearning For Llms, by Sungmin Cha et al.
Towards Robust and Parameter-Efficient Knowledge Unlearning for LLMs
by Sungmin Cha, Sungjun Cho, Dasol Hwang, Moontae Lee
First submitted to arxiv on: 13 Aug 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: 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 The proposed paper tackles the issue of unlearning sensitive data from Large Language Models (LLMs) without retraining them from scratch. This is crucial due to the risk of privacy and copyright violations when pretraining LLMs on massive textual corpora. The authors identify two major challenges: unstable optimization and catastrophic forgetting of relearned knowledge. They introduce two novel techniques to address these issues, including Inverted Hinge Loss and a data-adaptive initialization for LoRA adapters via low-rank approximation weighted with relative Fisher information. These methods allow for efficient and robust unlearning while maintaining the LLM’s reasoning and generative capabilities. The authors demonstrate their approach on various datasets and models, including GPT-Neo and Phi-1.5B. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper explores ways to remove sensitive data from Large Language Models (LLMs) without retraining them. This is important because LLMs can learn things they shouldn’t be learning. The authors found that a common method for unlearning, called Gradient Ascent, isn’t very good because it makes the model forget what it learned before. They then developed two new ways to unlearn: one uses a special loss function, and the other adjusts how the model starts out its learning process. These methods work well on different datasets and models. |
Keywords
» Artificial intelligence » Gpt » Hinge loss » Lora » Loss function » Optimization » Pretraining
