Title Ablation study of domain-specific focal layer for galaxy image classification with convolutional neural networks
Authors Mūžas, Tomas ; Misiukas Misiūnas, Andrius Vytautas ; Meškauskas, Tadas
DOI 10.22364/bjmc.2026.14.3.09
Full Text Download
Is Part of Baltic journal of modern computing.. Riga : University of Latvia. 2026, vol. 14, no. 3, p. 684-697.. ISSN 2255-8942. eISSN 2255-8950
Keywords [eng] Galaxy morphology classification ; Galaxy Zoo ; ablation study ; convolutional neural networks ; domain-specific layers
Abstract [eng] The aim of this research was to perform an ablation study of the domain-specific Focal convolution layer, proposed in our previous study. The study focused primarily on the images from ground telescopes -- Sloan Foundation and DECam, though Hubble space telescope data was explored as well. The applicability of the Focal layer was studied in two aspects. The first one was to verify whether a regular convolution can be replaced with the Focal to improve results on existing architectures. It was evaluated using two architectures of other authors, each having two variations: Original and Focal (all convolutions are Focal). The second aspect was to evaluate the impact of the Focal layer in isolation on our two best architectures from the previous study, each having three variations: NoFocal (all convolutions are regular), Original, and Focal. For ground-based data, the Focal variations of both other author architectures have achieved a statistically significant improvement of F1 scores compared to the Original variations. However, for both of our architectures, the best results were demonstrated by the Original variations -- a mix of both regular and Focal layers -- indicating that the layer should be applied conservatively. Additionally, based on the Original variations of both of our architectures, the layer has the highest impact when placed in the initial layers of the network, whereas its application in deeper layers should be avoided. As for the Hubble space telescope data, the Focal layer impact was minimal, what may have been caused by the absence of atmospheric turbulence or background objects.
Published Riga : University of Latvia
Type Journal article
Language English
Publication date 2026
CC license CC license description