Budget-Aware Data Selection for Low-Shot Image Classification: A Controlled Empirical Benchmark

DOI

10.2991/978-94-6239-828-3_5How to use a DOI?

Keywords

low-shot learning; budget-aware selection; coreset selection pseudo-label selection; RL-inspired selection; data efficiency

Abstract

This paper presents a controlled empirical benchmark of data selection for low-shot image classification under a limited annotation budget. Given a small labelled training set and an additional unlabelled candidate pool, we investigate which samples should be selected to improve downstream classification performance. We evaluate ten experimental conditions on CIFAR-10 and Fashion-MNIST using identical data splits, model architectures, training settings, budget levels, and random seeds. The evaluated conditions include random selection, pseudo-label- balanced selection, informativeness heuristics, geometry-based methods, augmentation, and an RL-inspired selector.

The results show that no single strategy performs best across both datasets. Augmentation achieves the highest aggregate area under the budget accuracy curve on CIFAR-10. On Fashion-MNIST, k-center achieves the highest value, although uniform random selection is only slightly behind. The RL-inspired selector is competitive on Fashion- MNIST and outperforms pseudo-label-balanced selection in several experimental settings. However, this improvement does not occur consistently across datasets and budgets. Analysis of the selected subsets also shows that high feature diversity alone does not guarantee better classification performance. Distributional alignment and stable class coverage remain important when the selection budget is limited. The main contribution of this work is therefore a reproducible empirical benchmark and an analysis of a hand-weighted RL-inspired selector, rather than a new reinforcement learning algorithm.

Copyright

© 2026 The Author(s)

Open Access

Open Access This chapter is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International License (http://creativecommons.org/licenses/by-nc/4.0/), which permits any noncommercial use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if changes were made.

Download article (PDF)

TY  - CONF
AU  - Aulia Ilmi Maghfira Ridwan
AU  - Zidane Putra Sanjaya
AU  - Bagus Prasetiyo
AU  - Novanto Yudistira
PY  - 2026
DA  - 2026/11/04
TI  - Budget-Aware Data Selection for Low-Shot Image Classification: A Controlled Empirical Benchmark
BT  - Proceedings of the 2026 11th International Conference on Sustainable Information Engineering and Technology (SIET 2026)
PB  - Atlantis Press
SP  - 43
EP  - 53
SN  - 2589-4900
UR  - https://doi.org/10.2991/978-94-6239-828-3_5
DO  - 10.2991/978-94-6239-828-3_5
ID  - Ridwan2026
ER  -