Summary of Finercut: Finer-grained Interpretable Layer Pruning For Large Language Models, by Yang Zhang et al.
FinerCut: Finer-grained Interpretable Layer Pruning for Large Language Models
by Yang Zhang, Yawei Li, Xinpeng Wang, Qianli Shen, Barbara Plank, Bernd Bischl, Mina Rezaei, Kenji Kawaguchi
First submitted to arxiv on: 28 May 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 FinerCut, a novel method for fine-grained layer pruning in transformer networks, aiming to reduce the environmental impact and computational requirements of Large Language Models (LLMs). By considering all self-attention and feed-forward network (FFN) layers within blocks as individual pruning candidates, FinerCut prunes layers with minimal output alteration. The approach is tested across 9 benchmarks, demonstrating retained performance of up to 95% while removing 30% of layers. Interestingly, the paper reveals that 42% of self-attention layers in Llama3-70B can be removed without compromising performance. FinerCut also provides insights into pruning behaviors, showing a preference for pruning self-attention layers at deeper consecutive decoder layers. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The researchers created a new way to make Large Language Models smaller and more efficient. They called it FinerCut. This method looks at each small part of the model separately and removes parts that don’t affect the model’s output much. The team tested FinerCut on many different tasks and found that it can keep the model’s performance almost the same, even when removing a lot of parts! They also discovered some interesting patterns about which parts get removed and why. |
Keywords
» Artificial intelligence » Decoder » Pruning » Self attention » Transformer
