An empirical study of task and feature correlations in the reuse of pre-trained models

Jama Hussein Mohamud · 2025

Read An empirical study of task and feature correlations in the reuse of pre-trained models by Jama Hussein Mohamud (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-01975.txt.

Description

Jama Hussein Mohamud and Willie Brink test why reusing parts of a trained neural network often works, in a paper from 2025. A common story is that one person trains a model, and another reuses its inner layers for a new task. Success is often credited to shared traits between the tasks.

Their tests show luck can play a large part. Gains grow as the two tasks align more, yet gains can appear even when tasks and inputs do not align at all. The choice of network and training method helps on its own. Reusing only low layers works best when tasks differ, and the best depth hints at how much tasks align.

Catalog record

CREATORJama Hussein Mohamud
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE40 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-01975

Browse the full library · About the Tower of Babel