Saved in:
Bibliographic Details
Main Authors: Chekalina, Viktoriia, Moskovskiy, Daniil, Cherniuk, Daria, Kurkin, Maxim, Kuznetsov, Andrey, Frolov, Evgeny
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2505.17974
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910964509573120
author Chekalina, Viktoriia
Moskovskiy, Daniil
Cherniuk, Daria
Kurkin, Maxim
Kuznetsov, Andrey
Frolov, Evgeny
author_facet Chekalina, Viktoriia
Moskovskiy, Daniil
Cherniuk, Daria
Kurkin, Maxim
Kuznetsov, Andrey
Frolov, Evgeny
contents The Fisher information is a fundamental concept for characterizing the sensitivity of parameters in neural networks. However, leveraging the full observed Fisher information is too expensive for large models, so most methods rely on simple diagonal approximations. While efficient, this approach ignores parameter correlations, often resulting in reduced performance on downstream tasks. In this work, we mitigate these limitations and propose Generalized Fisher-Weighted SVD (GFWSVD), a post-training LLM compression technique that accounts for both diagonal and off-diagonal elements of the Fisher information matrix, providing a more accurate reflection of parameter importance. To make the method tractable, we introduce a scalable adaptation of the Kronecker-factored approximation algorithm for the observed Fisher information. We demonstrate the effectiveness of our method on LLM compression, showing improvements over existing compression baselines. For example, at a 20 compression rate on the MMLU benchmark, our method outperforms FWSVD, which is based on a diagonal approximation of the Fisher information, by 5 percent, SVD-LLM by 3 percent, and ASVD by 6 percent compression rate.
format Preprint
id arxiv_https___arxiv_org_abs_2505_17974
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Generalized Fisher-Weighted SVD: Scalable Kronecker-Factored Fisher Approximation for Compressing Large Language Models
Chekalina, Viktoriia
Moskovskiy, Daniil
Cherniuk, Daria
Kurkin, Maxim
Kuznetsov, Andrey
Frolov, Evgeny
Machine Learning
Artificial Intelligence
The Fisher information is a fundamental concept for characterizing the sensitivity of parameters in neural networks. However, leveraging the full observed Fisher information is too expensive for large models, so most methods rely on simple diagonal approximations. While efficient, this approach ignores parameter correlations, often resulting in reduced performance on downstream tasks. In this work, we mitigate these limitations and propose Generalized Fisher-Weighted SVD (GFWSVD), a post-training LLM compression technique that accounts for both diagonal and off-diagonal elements of the Fisher information matrix, providing a more accurate reflection of parameter importance. To make the method tractable, we introduce a scalable adaptation of the Kronecker-factored approximation algorithm for the observed Fisher information. We demonstrate the effectiveness of our method on LLM compression, showing improvements over existing compression baselines. For example, at a 20 compression rate on the MMLU benchmark, our method outperforms FWSVD, which is based on a diagonal approximation of the Fisher information, by 5 percent, SVD-LLM by 3 percent, and ASVD by 6 percent compression rate.
title Generalized Fisher-Weighted SVD: Scalable Kronecker-Factored Fisher Approximation for Compressing Large Language Models
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2505.17974