Exploration of Random Attention Distribution in Counterfactual Attention Learning for Fine-Grained Visual Classification
JOIV : International Journal on Informatics Visualization · 31.05.2026 · 10 (3) · 1006
DOI: 10.62527/joiv.10.3.3850 ↗Kurz gefasst
Fachartikel in der Zeitschrift JOIV : International Journal on Informatics Visualization, online erschienen am 31.05.2026. Verfasst von Kanz Muhammad Hanif und 3 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY-SA 4.0).
Abstract (Originalsprache)
Attention is fundamental to perception and cognition. It is the mechanism by which we can selectively attend to relevant stimuli and make sense of them. This capability led to the emergence of machine learning paradigms that imitate the focus of human attention to improve performance on various computer vision tasks. Many existing models rely on weak supervision, which provides only the probability of a tendency. However, Counterfactual Attention Learning (CAL) addresses issues by using counterfactual reasoning to inspect the difference between the model’s attention and random attention, thereby lowering bias in training. Random attention is modeled using five distributions: uniform, normal, log-normal, exponential, and Bernoulli. It explores each distribution’s performance and identifies the best-fitting one to improve the accuracy of the CAL model's results on Fine-Grained Visual Classification (FGVC) tasks. The normal distribution outperforms the initial uniform distribution, yielding 0.44% and 0.03% higher accuracies on the FGVC-Aircraft and CUB-200-2011 datasets, respectively. Effective attention distributions in the CAL model show a balance between focusing on critical areas and maintaining context, smooth gradients that highlight relevant information without bias, and stable identification of crucial features. Thus, the normal distribution serves as a well-centered attention mechanism, improving the CAL model in FGVC tasks. However, since the CAL model relies on inherently random attention distributions, further research is needed to reduce stochastic variability and develop more stable, deterministic attention mechanisms for broader, more reliable applications.
Zitieren
APA
Hanif, K. M., Wibowo, S. A., Rahmania, R., & Usman, K. (2026). Exploration of Random Attention Distribution in Counterfactual Attention Learning for Fine-Grained Visual Classification. JOIV : International Journal on Informatics Visualization, 10(3), 1006. https://doi.org/10.62527/joiv.10.3.3850
Vancouver
Hanif KM, Wibowo SA, Rahmania R, Usman K. Exploration of Random Attention Distribution in Counterfactual Attention Learning for Fine-Grained Visual Classification. JOIV : International Journal on Informatics Visualization. 2026;10(3):1006. doi:10.62527/joiv.10.3.3850
BibTeX
@article{Hanif2026,
author = {Hanif, Kanz Muhammad and Wibowo, Suryo Adhi and Rahmania, Rissa and Usman, Koredianto},
title = {Exploration of Random Attention Distribution in Counterfactual Attention Learning for Fine-Grained Visual Classification},
journal = {JOIV : International Journal on Informatics Visualization},
year = {2026},
volume = {10},
number = {3},
pages = {1006},
doi = {10.62527/joiv.10.3.3850}
}
EUVIMED