AN INVESTIGATION OF CONVOLUTIONAL NEURAL NETWORK ARCHITECTURES ON THE FASHION-MNIST DATASET UNDER CONSTRAINED PARAMETER BUDGETS
Ukrainian Journal of Information Technology · 2026 · 8 (1) · 1-7
DOI: 10.23939/ujit2026.01.001 ↗Kurz gefasst
Fachartikel in der Zeitschrift Ukrainian Journal of Information Technology, 2026. Verfasst von O. Y. Beznosyk und 1 weiteren Autor:innen. Die Arbeit nennt 12 Literaturangaben.
Abstract (Originalsprache)
This paper addresses the problem of finding an effective convolutional neural network architecture under a constrained parameter budget. This problem is relevant for deployment on low-power devices without hardware acceleration, including IoT devices and microcontrollers. The problem was studied on the Fashion-MNIST dataset. Several architectures with varying layer counts and configurations were taken as candidates. The parameter budget was fixed at 50000 trainable parameters. One optimizer, namely SGD, was evaluated. Each architecture was trained for 20 epochs under a unified training protocol and the best-performing architecture was selected. That architecture was studied over a larger number of epochs and across multiple random seeds. Five architecture variants were compared such as narrow_deep, pyramid, inv_pyramid, fc_bottleneck, and wide_shallow. Different feature aggregation strategies were analyzed, specifically comparing Global Average Pooling (GAP-head) and fully connected classifiers with a bottleneck (FC-head). The experiments show that the narrow_deep architecture with channel configuration 16-24-28-36-40-48 achieved the highest test accuracy. However, the most computationally efficient results were achieved by the fc_bottleneck configuration, with approximately the same accuracy. It was also found that the random seed can significantly affect training outcomes and, in some cases, cause accuracy degradation of up to 10% regardless of the number of epochs. The findings allow practitioners to select an effective layer count, layer width, and optimizer for building a CNN within a strict 50000-parameter budget on Fashion-MNIST. All experiments were conducted on CPU-only hardware (Intel Core i7, 32 GB RAM, Debian 12), which confirms the practical applicability of this approach under limited computational resources. This problem formulation and its results provide a foundation for further investigation of generalized and specialized models under strict parameter constraints.
Zitieren
APA
Beznosyk, O. Y., & Kharchenko, K. V. (2026). AN INVESTIGATION OF CONVOLUTIONAL NEURAL NETWORK ARCHITECTURES ON THE FASHION-MNIST DATASET UNDER CONSTRAINED PARAMETER BUDGETS. Ukrainian Journal of Information Technology, 8(1), 1-7. https://doi.org/10.23939/ujit2026.01.001
Vancouver
Beznosyk OY, Kharchenko KV. AN INVESTIGATION OF CONVOLUTIONAL NEURAL NETWORK ARCHITECTURES ON THE FASHION-MNIST DATASET UNDER CONSTRAINED PARAMETER BUDGETS. Ukrainian Journal of Information Technology. 2026;8(1):1-7. doi:10.23939/ujit2026.01.001
BibTeX
@article{Beznosyk2026,
author = {Beznosyk, O. Y. and Kharchenko, K. V.},
title = {AN INVESTIGATION OF CONVOLUTIONAL NEURAL NETWORK ARCHITECTURES ON THE FASHION-MNIST DATASET UNDER CONSTRAINED PARAMETER BUDGETS},
journal = {Ukrainian Journal of Information Technology},
year = {2026},
volume = {8},
number = {1},
pages = {1--7},
doi = {10.23939/ujit2026.01.001}
}
EUVIMED