Practical Galaxy Morphology Tools from Deep Supervised Representation Learning

“In fact, anymodern neural network trained for classification or regression can be used as a representation learning framework (see, for example, Walmsley et al. 2022a for an application to galaxy morphology).”

“, between training and target datasets) has a few implementations for galaxy morphology classifications ([101, 102, 103, 76]).”

“Consequently, a wide range of supervised and semi-supervised learning methods have been developed [5, 6].”

“Along these lines, deep representation learning models that encode more detailed characterizations of galaxies ( e.g., Walmsley et al. 2022; Omori et al. 2023; Eisert et al. 2024 ) may also provide more complete morphological representations that can serve as useful input into HaloFlow .”

“Work has been carried out by Walmsley et al. (2022a) that combines these visual classifications from Galaxy Zoo with machine learning, allowing for many more classifications, and also enabling researchers to locate anomalies within their datasets (Walmsley et al. 2022b).”

“Walmsley et al. [35] have demonstrated success across three tasks.”

“Similarity search has been applied in various astronomical studies, including galaxy morphology research, such as (Walmsley et al. 2022), which focuses on the Galaxy Zoo projects.”

“Each interacting system was found using the European Space Agency’s new platform ESA Datalabs, which allowed us to directly apply an advanced CNN— Zoobot—to the entire Hubble science archive.”

“Studies in other areas of astronomy have more broadly adopted pre-training and off-the-shelf architectures for their vision models (e.g., Wu & Boada 2019; Walmsley et al. 2022; Vafaei Sadr et al. 2023; Mo-hale & Lochner 2024).”

“This function, described in detail in Walmsley et al. (2022b), includes a hyperparameter ϵ which can be used to tune the trade-off between exploitation (low values of ϵ ) and exploration (high values of ϵ ).”

“Walmsley et al. [9] suggested that the representation learned by Zoobot may arrange galaxies in patterns beyond those required to predict the supervised labels.”

“Hence, to efficiently analyze these large datasets, it is critical to consider leveraging human-machine optimization frameworks, such as active learning strategies with human-in-the-loop ideologies [4, 9].”

“Particularly, a recent study has detected over 40,000 ring galaxies using a machine learning-based approach, highlighting the potential of computational methods for large-scale galaxy classification (Walmsley et al. 2022).”

“Several techniques [25], [37], [45], [21], [44] have leveraged the class labels to generate powerful representations, which has led to state-of-the-art performance.”

“A detailed description and methodology of Zoobot can be found in M. Walmsley et al. ( 2022b, 2022c, 2023 ) .”

“, [12, 17]) show lower gains than image segmentation based tasks (e.”

“To ensure the reliability and representativeness of our data set, we used volunteer voting data in the Galaxy Zoo 2 (GZ2) project (M. Walmsley et al. 2022b).”

“Notably, similar feature extraction methods have shown potential in the field of astronomical image analysis (e.g. Etsebeth et al., 2023; Walmsley et al., 2022).”

“…C. J. Lintott et al. 2008; M. Banerji et al. 2010 ) and more complex tools ( S. Dieleman et al. 2015; H. Domínguez Sánchez et al. 2018; P. H. Barchi et al. 2020; R. Hausen & B. E. Robertson 2020; G. Martin et al. 2020; T.-Y. Cheng et al. 2021; C. Tohill et al. 2021; M. Walmsley et al. 2022 ) .”

“It is proven that a CNN backbone trained on astrophysical images outperform those trained on terrestrial images in astronomical object classifiers [73] and detectors [75].”

“Walmsley et al. (2022) pioneered the approach of using a pretrained network as a feature extractor for optical Galaxy Zoo images, showing that an algorithm trained on a complex classification task creates a representation of the galaxy images which can be used for downstream tasks, including anomaly…”

“…predictor transfers unchanged between surveys, consistent with domain-adaptation work showing explicit alignment objectives, adapters, and target-domain fine-tuning improve transfer between simulations and observations and across surveys [ ´Ciprijanovi´c et al., 2021, 2023, Walmsley et al., 2022].”

“…image representations can support semantic similarity search and light-weight supervised probes over large unlabeled collections ( G. Stein et al. 2021, 2022 ) , while supervised representation learning can provide reusable features for several galaxy-morphology tasks ( M. Walmsley et al. 2022a ) .”

“17 This study synthesises community perspectives obtained through surveys, workshops, and stakeholder engagement activities.”

“Furthermore, M. Walmsley et al. ( 2022a ) explored a hybrid contrastive learning method that adapts the BYOL framework and incorporates a Dirichlet loss to leverage both massive unlabeled images and existing ( and noisy ) volunteer labels, achieving significant performance gains, particularly in…”

“…their limited sample size coupled with high-dimensional features, these general models often exhibit suboptimal recognition accuracy and insufficient generalization ability, rendering them unable to meet the requirements for precise selection and study ( Angora et al. 2020; Walmsley et al. 2022 ) .”

“…have demonstrated the ability to learn meaningful representations of astronomical phenomena that transfer effectively to tasks beyond their original classification objective, including anomaly detection and morphological analysis (Walmsley et al., 2022; Etsebeth et al., 2023; Gupta et al., 2025).”

“Darg et al. 2010a,b), ring galaxies (M. Walmsley et al. 2022), and dust lane galaxies (e.g., the K12 sample).”

“In response to the challenges posed by the vast data volumes generated by next-generation radio telescopes, machine learning has emerged as a key analytical tool (e.g., [10]–[27]).”

“For instance, content-based image retrieval (CBIR) [2, 3], and text-based image retrieval (TBIR) [4] approaches have been applied in finding the ‘nearest neighbour’ search objects from a reference database in astronomy.”