Saved in:
Bibliographic Details
Main Authors: Yu, Lei, Niu, Jingcheng, Zhu, Zining, Chen, Xi, Penn, Gerald
Format: Preprint
Published: 2024
Subjects:
Online Access:https://arxiv.org/abs/2407.03779
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912613773869056
author Yu, Lei
Niu, Jingcheng
Zhu, Zining
Chen, Xi
Penn, Gerald
author_facet Yu, Lei
Niu, Jingcheng
Zhu, Zining
Chen, Xi
Penn, Gerald
contents In this paper, we introduce DiscoGP, a novel framework for extracting self-contained modular units, or sheaves, within neural language models (LMs). Sheaves extend the concept of functional circuits, a unit widely explored in interpretability research, by considering not only subsets of edges in an LM's computation graph but also the model's weight parameters. Our framework identifies sheaves through a gradient-based pruning algorithm that operates on both of these in such a way that reduces the original LM to a sparse skeleton that preserves certain core capabilities. Experimental results demonstrate that, across a range of linguistic and reasoning tasks, DiscoGP extracts sheaves that preserve 93%-100% of a model's performance on the identified task while comprising only 1%-7% of the original weights and connections. Furthermore, our analysis reveals that, compared to previously identified LM circuits, the sheaves discovered by DiscoGP exhibit superior modularity and functional fidelity. Extending our method to the neuron level also unveils novel insights into the inner workings of LLMs
format Preprint
id arxiv_https___arxiv_org_abs_2407_03779
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Sheaf Discovery with Joint Computation Graph Pruning and Flexible Granularity
Yu, Lei
Niu, Jingcheng
Zhu, Zining
Chen, Xi
Penn, Gerald
Computation and Language
Machine Learning
In this paper, we introduce DiscoGP, a novel framework for extracting self-contained modular units, or sheaves, within neural language models (LMs). Sheaves extend the concept of functional circuits, a unit widely explored in interpretability research, by considering not only subsets of edges in an LM's computation graph but also the model's weight parameters. Our framework identifies sheaves through a gradient-based pruning algorithm that operates on both of these in such a way that reduces the original LM to a sparse skeleton that preserves certain core capabilities. Experimental results demonstrate that, across a range of linguistic and reasoning tasks, DiscoGP extracts sheaves that preserve 93%-100% of a model's performance on the identified task while comprising only 1%-7% of the original weights and connections. Furthermore, our analysis reveals that, compared to previously identified LM circuits, the sheaves discovered by DiscoGP exhibit superior modularity and functional fidelity. Extending our method to the neuron level also unveils novel insights into the inner workings of LLMs
title Sheaf Discovery with Joint Computation Graph Pruning and Flexible Granularity
topic Computation and Language
Machine Learning
url https://arxiv.org/abs/2407.03779