Summary of An Efficient Rehearsal Scheme For Catastrophic Forgetting Mitigation During Multi-stage Fine-tuning, by Andrew Bai et al.
An Efficient Rehearsal Scheme for Catastrophic Forgetting Mitigation during Multi-stage Fine-tuning
by Andrew Bai, Chih-Kuan Yeh, Cho-Jui Hsieh, Ankur Taly
First submitted to arxiv on: 12 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 novel approach to alleviate the “catastrophic forgetting” of prior knowledge during fine-tuning of foundational models in NLP. The authors focus on a fixed computational budget instead of a fixed memory buffer, introducing the mix-cd scheme that prioritizes rehearsal of “collateral damage” samples. These are correct predictions by the prior model but forgotten by the incrementally tuned one. The method efficiently estimates density without additional model inferences, outperforming leading continual learning methods in compute-constrained settings. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper helps us remember things we learned before and forget less when we learn new things. It’s like how our brains work! They introduce a way to use the same amount of computer power but still keep what we knew before. This is important because it makes machines smarter without making them too slow or expensive. |
Keywords
* Artificial intelligence * Continual learning * Fine tuning * Nlp
