Task Singular Vectors: Reducing Task Interference in Model Merging
Fuente:
arXiv
Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2024
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866917976457871360 |
|---|---|
| author | Gargiulo, Antonio Andrea Crisostomi, Donato Bucarelli, Maria Sofia Scardapane, Simone Silvestri, Fabrizio Rodolà, Emanuele |
| author_facet | Gargiulo, Antonio Andrea Crisostomi, Donato Bucarelli, Maria Sofia Scardapane, Simone Silvestri, Fabrizio Rodolà, Emanuele |
| contents | Task Arithmetic has emerged as a simple yet effective method to merge models without additional training. However, by treating entire networks as flat parameter vectors, it overlooks key structural information and is susceptible to task interference. In this paper, we study task vectors at the layer level, focusing on task layer matrices and their singular value decomposition. In particular, we concentrate on the resulting singular vectors, which we refer to as Task Singular Vectors (TSV). Recognizing that layer task matrices are often low-rank, we propose TSV-Compress (TSV-C), a simple procedure that compresses them to 10% of their original size while retaining 99% of accuracy. We further leverage this low-rank space to define a new measure of task interference based on the interaction of singular vectors from different tasks. Building on these findings, we introduce TSV-Merge (TSV-M), a novel model merging approach that combines compression with interference reduction, significantly outperforming existing methods. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2412_00081 |
| institution | arXiv |
| publishDate | 2024 |
| record_format | arxiv |
| spellingShingle | Task Singular Vectors: Reducing Task Interference in Model Merging Gargiulo, Antonio Andrea Crisostomi, Donato Bucarelli, Maria Sofia Scardapane, Simone Silvestri, Fabrizio Rodolà, Emanuele Machine Learning I.5.1; I.4.2; I.2.10 Task Arithmetic has emerged as a simple yet effective method to merge models without additional training. However, by treating entire networks as flat parameter vectors, it overlooks key structural information and is susceptible to task interference. In this paper, we study task vectors at the layer level, focusing on task layer matrices and their singular value decomposition. In particular, we concentrate on the resulting singular vectors, which we refer to as Task Singular Vectors (TSV). Recognizing that layer task matrices are often low-rank, we propose TSV-Compress (TSV-C), a simple procedure that compresses them to 10% of their original size while retaining 99% of accuracy. We further leverage this low-rank space to define a new measure of task interference based on the interaction of singular vectors from different tasks. Building on these findings, we introduce TSV-Merge (TSV-M), a novel model merging approach that combines compression with interference reduction, significantly outperforming existing methods. |
| title | Task Singular Vectors: Reducing Task Interference in Model Merging |
| topic | Machine Learning I.5.1; I.4.2; I.2.10 |
| url | https://arxiv.org/abs/2412.00081 |