Dissecting Large Language Models

Abstract

Understanding and shaping the behaviour of Large Language Models (LLMs) is increasingly important as applications become more powerful and more frequently adopted. This paper introduces a machine unlearning method specifically designed for LLMs. We introduce a selective pruning method for LLMs that removes neurons based on their relative importance on a targeted capability compared to overall network performance. This approach is a compute- and data-efficient method for identifying and removing neurons that enable specific behaviours. Our findings reveal that both feed-forward and attention neurons in LLMs are specialized; that is, for specific tasks, certain neurons are more crucial than others.

Cite

Text

Pochinkov and Schoots. "Dissecting Large Language Models." NeurIPS 2023 Workshops: SoLaR, 2023.

Markdown

[Pochinkov and Schoots. "Dissecting Large Language Models." NeurIPS 2023 Workshops: SoLaR, 2023.](https://mlanthology.org/neuripsw/2023/pochinkov2023neuripsw-dissecting/)

BibTeX

@inproceedings{pochinkov2023neuripsw-dissecting,
  title     = {{Dissecting Large Language Models}},
  author    = {Pochinkov, Nicky and Schoots, Nandi},
  booktitle = {NeurIPS 2023 Workshops: SoLaR},
  year      = {2023},
  url       = {https://mlanthology.org/neuripsw/2023/pochinkov2023neuripsw-dissecting/}
}