Forgetting of task-specific knowledge in model merging-based continual learning

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Hess, Timm, van de Ven, Gido M, Tuytelaars, Tinne
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912511948750848
author Hess, Timm
van de Ven, Gido M
Tuytelaars, Tinne
author_facet Hess, Timm
van de Ven, Gido M
Tuytelaars, Tinne
contents This paper investigates the linear merging of models in the context of continual learning (CL). Using controlled visual cues in computer vision experiments, we demonstrate that merging largely preserves or enhances shared knowledge, while unshared task-specific knowledge rapidly degrades. We further find that merging models from an incremental training process consistently outperforms merging models trained in parallel.
format Preprint
id arxiv_https___arxiv_org_abs_2507_23311
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Forgetting of task-specific knowledge in model merging-based continual learning
Hess, Timm
van de Ven, Gido M
Tuytelaars, Tinne
Computer Vision and Pattern Recognition
This paper investigates the linear merging of models in the context of continual learning (CL). Using controlled visual cues in computer vision experiments, we demonstrate that merging largely preserves or enhances shared knowledge, while unshared task-specific knowledge rapidly degrades. We further find that merging models from an incremental training process consistently outperforms merging models trained in parallel.
title Forgetting of task-specific knowledge in model merging-based continual learning
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2507.23311