Dissecting Language Models: Machine Unlearning via Selective Pruning

Nicky Pochinkov, Nandi Schoots

Primary Area: visualization or interpretation of learned representations

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: language models, pruning, machine unlearning, capability removal, intepretability, modularity

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: We propose a method for machine unlearning via pruning, and find that language models have task specialized neurons.

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.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 5824