Can a Short Training Run Choose a Generative Data Augmentation?

Image recognition models need diverse datasets, but collecting such images takes a lot of cost and time. For this reason, images generated by AI have recently been used. However, adding generated images sometimes raises recognition accuracy and sometimes lowers it. To find out which generated images have a good effect on the model, the only way is to run the training to the end, which is costly. In this work, we study whether, in image classification with little data, we can predict which data will have a good effect before running the full training. Evaluation metrics for generated images such as FID, the conventional approach, were not useful for predicting whether the final accuracy of the model improves. For example, images generated in a pencil-sketch style were judged by FID to be the farthest from the real images, yet they raised accuracy the most. Conversely, when we replaced the generated images with ones labelled with a class different from their content (e.g., a dog image labelled ``cat''), accuracy dropped sharply, but FID barely changed. On the other hand, when we run a short training of about 6\% of the full training and, at its end, compare how well the model predicts real and generated images, the better the model predicts the generated images compared with the real images, the higher the full-training accuracy was. With this trend, we could predict the accuracy ranking of the candidates fairly well. If we use this trend and run the full training only for the top three, we could pick a nearly best candidate at 30--70\% of the cost of training all of them. However, it is weak at picking the best one among candidates that already have high accuracy, and it is fooled by images that are ``easy to learn but not useful,'' such as duplicated images.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-12
DOI
https://doi.org/10.5281/zenodo.24156590
Primary Topic
Domain Adaptation and Few-Shot Learning
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Can a Short Training Run Choose a Generative Data Augmentation?

Ikumi Taue
Zenodo (CERN European Organization for Nuclear Research)
Domain Adaptation and Few-Shot Learning
preprint

Can a Short Training Run Choose a Generative Data Augmentation?

Ikumi Taue
preprint en

Abstract

Image recognition models need diverse datasets, but collecting such images takes a lot of cost and time. For this reason, images generated by AI have recently been used. However, adding generated images sometimes raises recognition accuracy and sometimes lowers it. To find out which generated images have a good effect on the model, the only way is to run the training to the end, which is costly. In this work, we study whether, in image classification with little data, we can predict which data will have a good effect before running the full training. Evaluation metrics for generated images such as FID, the conventional approach, were not useful for predicting whether the final accuracy of the model improves. For example, images generated in a pencil-sketch style were judged by FID to be the farthest from the real images, yet they raised accuracy the most. Conversely, when we replaced the generated images with ones labelled with a class different from their content (e.g., a dog image labelled ``cat''), accuracy dropped sharply, but FID barely changed. On the other hand, when we run a short training of about 6\% of the full training and, at its end, compare how well the model predicts real and generated images, the better the model predicts the generated images compared with the real images, the higher the full-training accuracy was. With this trend, we could predict the accuracy ranking of the candidates fairly well. If we use this trend and run the full training only for the top three, we could pick a nearly best candidate at 30--70\% of the cost of training all of them. However, it is weak at picking the best one among candidates that already have high accuracy, and it is fooled by images that are ``easy to learn but not useful,'' such as duplicated images.

Zenodo (CERN European Organization for Nuclear Research)
Osaka Institute of Technology (JP)
Domain Adaptation and Few-Shot Learning
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon