Class Of Galaxies Research Articles

Optically active and optically inactive radio galaxies as sub-populations of the main galaxy sample of the SDSS

We use the ROGUE I and II catalogues of radio sources associated with optical galaxies to revisit the characterization of radio active galactic nuclei (AGNs) in terms of radio luminosities and properties derived from the analyses of the optical spectra of their associated galaxies. We propose a physically based classification of radio galaxies into `optically inactive' and `optically active' (OPARGs and OPIRGs). In our sample, there are 14082 OPIRGs and 2721 OPARGs. After correcting for the Malmquist bias, we compared the global properties of our two classes of radio galaxies and put them in the context of the global population of galaxies. To compare the Eddington ratios of OPARGs with those of Seyferts, we devised a method to obtain the bolometric luminosities of these objects, taking into account the contribution of young stars to the observed line emission. We provide formulae to derive bolometric luminosities from the luminosity. We find that the distributions of radio luminosities of OPARGs and OPIRGs are undistinguishable. On average, the black hole masses and stellar masses in OPIRGs are larger than in OPARGs. OPARGs show signs of some recent star formation. Plotting the OPARGs in the BPT diagram and comparing their distribution with that of the remaining galaxies, we find that there is a sub-family of very high excitation OPARGs at the top of the AGN wing. This group is slightly displaced towards the left of the rest of the AGN galaxies, suggesting a stronger ionizing radiation field with respect to the gas pressure. Only very-high excitation radio galaxies (VHERGs) have Eddington ratios higher than $10^ $, which are canonically considered as the lower limit for the occurrence of radiative efficient accretion. If our estimates of the bolometric luminosities are correct, this means than only a small proportion of mainstream HERGs are indeed radiatively efficient.

A hidden active galactic nucleus population: the first radio luminosity functions constructed by physical process

ABSTRACT Both star formation (SF) and active galactic nuclei (AGNs) play an important role in galaxy evolution. Statistically quantifying their relative importance can be done using radio luminosity functions (RLFs). Until now these relied on galaxy classifications, where sources with a mixture of radio emission from SF and AGN are labelled as either a star-forming galaxy or an AGN. This can cause the misestimation of the relevance of AGN. Brightness temperature measurements at 144 MHz with the International LOw Frequency ARray telescope can separate radio emission from AGN and SF. We use the combination of sub-arcsec and arcsec resolution imaging of 7497 sources in the Lockman Hole and ELAIS-N1 fields to identify AGN components in the sub-arcsec resolution images and subtract them from the total flux density, leaving flux density from SF only. We construct, for the first time, RLFs by physical process, either SF or AGN activity, revealing a hidden AGN population at $L_{\textrm {144 MHz}}$$\lt 10^{24}$ W Hz$^{-1}$. This population is 1.56 $\pm$ 0.06 more than expected for $0.5\lt z\lt 2.0$ when comparing to RLFs by galaxy classification. The star-forming population has only 0.90 $\pm$ 0.02 of the expected SF. These ‘hidden’ AGNs can have significant implications for the cosmic SF rate and kinetic luminosity densities.

At First Sight! Zero-shot Classification of Astronomical Images with Large Multimodal Models

Abstract Vision-Language multimodal Models (VLMs) offer the possibility for zero-shot classification in astronomy: i.e., classification via natural language prompts, with no training. We investigate two models, GPT-4o and LLaVA-NeXT, for zero-shot classification of low-surface brightness galaxies and artifacts, as well as morphological classification of galaxies. We show that with natural language prompts these models achieved significant accuracy (above 80% typically) without additional training/fine tuning. We discuss areas that require improvement, especially for LLaVA-NeXT, which is an open source model. Our findings aim to motivate the astronomical community to consider VLMs as a powerful tool for both research and pedagogy, with the prospect that future custom-built or fine-tuned models could perform better.

MEGS: Morphological Evaluation of Galactic Structur. Principal component analysis as a galaxy morphology model

Understanding the morphology of galaxies is a critical aspect of astrophysics research, providing insight into the formation, evolution, and physical properties of these vast cosmic structures. Various observational and computational methods have been developed to quantify galaxy morphology, and with the advent of large galaxy simulations, the need for automated and effective classification methods has become increasingly important. This paper investigates the use of principal component analysis (PCA) as an interpretable dimensionality reduction algorithm for galaxy morphology using the IllustrisTNG cosmological simulation dataset with the aim of developing a generative model for galaxies. We first generate a dataset of 2D images and 3D cubes of galaxies from the IllustrisTNG simulation, focusing on the mass, metallicity, and stellar age distribution of each galaxy. PCA is then applied to this data, transforming it into a lower-dimensional image space, where closeness of data points corresponds to morphological similarity. We find that PCA can effectively capture the key morphological features of galaxies, with a significant proportion of the variance in the data being explained by a small number of components. With our method we achieve a dimensionality reduction by a factor of $ for 2D images and $ for 3D cubes at a reconstruction accuracy below five percent. Our results illustrate the potential of PCA in compressing large cosmological simulations into an interpretable generative model for galaxies that can easily be used in various downstreaming tasks such as galaxy classification and analysis.

Galaxy Classification Using EWGC

The Enhanced Wide-field Galaxy Classification Network (EWGC) is a novel architecture designed to classify spiral and elliptical galaxies using Wide-field Infrared Survey Explorer (WISE) images. The EWGC achieves an impressive classification accuracy of 90.02%, significantly outperforming the previously developed WGC network and underscoring its superior performance in galaxy morphology classification. Remarkably, the network demonstrates a consistent accuracy of 90.02% when processing both multi-target and single-target images. Such robustness indicates the EWGC’s versatility and potential for various applications in galaxy classification tasks.

Retrieval of the physical parameters of galaxies from WEAVE-StePS-like data using machine learning

Context. The William Herschel Telescope Enhanced Area Velocity Explorer (WEAVE) is a new, massively multiplexing spectrograph that allows us to collect about one thousand spectra over a 3 square degree field in one observation. The WEAVE Stellar Population Survey (WEAVE-StePS) in the next 5 years will exploit this new instrument to obtain high-S/N spectra for a magnitude-limited (IAB = 20.5) sample of ∼25 000 galaxies at moderate redshifts (z ≥ 0.3), providing insights into galaxy evolution in this as yet unexplored redshift range. Aims. We aim to test novel techniques for retrieving the key physical parameters of galaxies from WEAVE-StePS spectra using both photometric and spectroscopic (spectral indices) information for a range of noise levels and redshift values. Methods. We simulated ∼105 000 galaxy spectra assuming star formation histories with an exponentially declining star formation rate, covering a wide range of ages, stellar metallicities, specific star formation rates (sSFRs), and dust extinction values. We considered three redshifts (i.e. z = 0.3, 0.55, and 0.7), covering the redshift range that WEAVE-StePS will observe. We then evaluated the ability of the random forest and K-nearest neighbour algorithms to correctly predict the average age, metallicity, sSFR, dust attenuation, and time since the bulk of formation, assuming no measurement errors. We also checked how much the predictive ability deteriorates for different noise levels, with S/NI,obs = 10, 20, and 30, and at different redshifts. Finally, the retrieved sSFR was used to classify galaxies as part of the blue cloud, green valley, or red sequence. Results. We find that both the random forest and K-nearest neighbour algorithms accurately estimate the mass-weighted ages, u-band-weighted ages, and metallicities with low bias. The dispersion varies from 0.08–0.16 dex for age and 0.11–0.25 dex for metallicity, depending on the redshift and noise level. For dust attenuation, we find a similarly low bias and dispersion. For the sSFR, we find a very good constraining power for star-forming galaxies, log sSFR ≳ −11, where the bias is ∼0.01 dex and the dispersion is ∼0.10 dex. However, for more quiescent galaxies, with log sSFR ≲ −11, we find a higher bias, ranging from 0.61 to 0.86 dex, and a higher dispersion, ∼0.4 dex, depending on the noise level and redshift. In general, we find that the random forest algorithm outperforms the K-nearest neighbours. Finally, we find that the classification of galaxies as members of the green valley is successful across the different redshifts and S/Ns. Conclusions. We demonstrate that machine learning algorithms can accurately estimate the physical parameters of simulated galaxies for a WEAVE-StePS-like dataset, even at relatively low S/NI, obs = 10 per Å spectra with available ancillary photometric information. A more traditional approach, Bayesian inference, yields comparable results. The main advantage of using a machine learning algorithm is that, once trained, it requires considerably less time than other methods.

Identifying Mergers in the Legacy Surveys with Few-shot Learning

Galaxy mergers exert a pivotal influence on the evolutionary trajectory of galaxies and the expansive development of cosmic structures. The primary challenge encountered in machine learning–based identification of merging galaxies arises from the scarcity of meticulously labeled data sets specifically dedicated to merging galaxies. In this paper, we propose a novel framework utilizing few-shot learning techniques to identify galaxy mergers in the Legacy Surveys. Few-shot learning enables effective classification of merging galaxies even when confronted with limited labeled training samples. We employ a deep convolutional neural network architecture trained on data sets sampled from Galaxy Zoo Decals to learn essential features and generalize to new instances. Our experimental results demonstrate the efficacy of our approach, achieving high accuracy and precision in identifying galaxy mergers with few labeled training samples. Furthermore, we investigate the impact of various factors, such as the number of training samples and network architectures, on the performance of the few-shot learning model. The proposed methodology offers a promising avenue for automating the identification of galaxy mergers in large-scale surveys, facilitating the comprehensive study of galaxy evolution and structure formation. In pursuit of identifying galaxy mergers, our methodology is applied to analyze the Data Release 9 of the Dark Energy Spectroscopic Instrument Legacy Imaging Surveys. As a result, we have unveiled an extensive catalog encompassing 648,183 galaxy merger candidates. We publicly release the catalog alongside this paper.

Integrating Fuzzy C-Means Clustering and Explainable AI for Robust Galaxy Classification

The classification of galaxies has significantly advanced using machine learning techniques, offering deeper insights into the universe. This study focuses on the typology of galaxies using data from the Galaxy Zoo project, where classifications are based on the opinions of non-expert volunteers, introducing a degree of uncertainty. The objective of this study is to integrate Fuzzy C-Means (FCM) clustering with explainability methods to achieve a precise and interpretable model for galaxy classification. We applied FCM to manage this uncertainty and group galaxies based on their morphological characteristics. Additionally, we used explainability methods, specifically SHAP (SHapley Additive exPlanations) values and LIME (Local Interpretable Model-Agnostic Explanations), to interpret and explain the key factors influencing the classification. The results show that using FCM allows for accurate classification while managing data uncertainty, with high precision values that meet the expectations of the study. Additionally, SHAP values and LIME provide a clear understanding of the most influential features in each cluster. This method enhances our classification and understanding of galaxies and is extendable to environmental studies on Earth, offering tools for environmental management and protection. The presented methodology highlights the importance of integrating FCM and XAI techniques to address complex problems with uncertain data.

Supervised star, galaxy, and QSO classification with sharpened dimensionality reduction

We explored the use of broadband colors to classify stars, galaxies, and quasi-stellar objects (QSOs). Specifically, we applied sharpened dimensionality reduction (SDR)-aided classification to this problem, with the aim of enhancing cluster separation in the projections of high-dimensional data clusters to allow for better classification performance and more informative projections. The main objective of this work was to apply SDR to large sets of broadband colors derived from the CPz catalog to obtain projections with clusters of star, galaxy, and QSO data that exhibit a high degree of separation. The SDR method achieves this by combining density-based clustering with conventional dimensionality-reduction techniques. To make SDR scalable and have the ability to project samples using the earlier-computed projection, we used a deep neural network trained to reproduce the SDR projections. Subsequently classification was done by applying a $k$-nearest neighbors ($k$-NN) classifier to the sharpened projections. Based on a qualitative and quantitative analysis of the embeddings produced by SDR, we find that SDR consistently produces accurate projections with a high degree of cluster separation. A number of projection performance metrics are used to evaluate this separation, including the trustworthiness, continuity, Shepard goodness, and distribution consistency metrics. Using the $k$-NN classifier and consolidating the results of various data sets, we obtain precisions of 99.7<!PCT!>, 98.9<!PCT!>, and 98.5<!PCT!> for classifying stars, galaxies, and QSOs, respectively. Furthermore, we achieve completenesses of 97.8<!PCT!>, 99.3<!PCT!>, and 86.8<!PCT!>, respectively. In addition to classification, we explore the structure of the embeddings produced by SDR by cross-matching with data from Gaia DR3, Galaxy Zoo 1, and a catalog of specific star formation rates, stellar masses, and dust luminosities. We discover that the embeddings reveal astrophysical information, which allows one to understand the structure of the high-dimensional broadband color data in greater detail. We find that SDR-aided star, galaxy, and QSO classification performs comparably to another unsupervised learning method using hierarchical density-based spatial clustering of applications with noise (HDBSCAN) but offers advantages in terms of scalability and interpretability. Furthermore, it outperforms traditional color selection methods in terms of QSO classification performance. Overall, we demonstrate the potential of SDR-aided classification to provide an accurate and physically insightful classification of astronomical objects based on their broadband colors.

Bayesian and convolutional networks for hierarchical morphological classification of galaxies

Bayesian and convolutional networks for hierarchical morphological classification of galaxies

FNet II: spectral classification of quasars, galaxies, stars, and broad absorption line (BAL) quasars

<tt>FNet II</tt>: spectral classification of quasars, galaxies, stars, and broad absorption line (BAL) quasars

Exploring galaxy properties of eCALIFA with contrastive learning

Contrastive learning (CL) has emerged as a potent tool for building meaningful latent representations of galaxy properties across a broad spectrum of wavelengths, ranging from optical and infrared to radio frequencies. These latent representations facilitate a variety of downstream tasks, including galaxy classification, similarity searches in extensive datasets, and parameter estimation, which is why they are often referred to as foundation models for galaxies. In this study, we employ CL on the latest extended data release from the Calar Alto Legacy Integral Field Area (CALIFA) survey, which encompasses a total of 895 galaxies with enhanced spatial resolution that reaches the limits imposed by natural seeing (FWHMPSF ∼ 1.5). We demonstrate that CL can be effectively applied to Integral Field Unit (IFU) surveys, even with relatively small training sets, to construct meaningful embedding where galaxies are well separated based on their physical properties. We discover that the strongest correlations in the embedding space are observed with the equivalent width of Hα, galaxy morphology, stellar metallicity, luminosity-weighted age, stellar surface mass density, the [NII]/Hα ratio, and stellar mass, in descending order of correlation strength. Additionally, we illustrate the feasibility of unsupervised separation of galaxy populations along the star formation main sequence, successfully identifying the blue cloud and the red sequence in a two-cluster scenario, and the green valley population in a three-cluster scenario. Our findings indicate that galaxy luminosity profiles have minimal impact on the construction of the embedding space, suggesting that morphology and spectral features play a more significant role in distinguishing between galaxy populations. Moreover, we explore the use of CL for detecting variations in galaxy population distributions across different large-scale structures, including voids, clusters, and filaments and walls. Nonetheless, we acknowledge the limitations of the CL framework and our specific training set in detecting subtle differences in galaxy properties, such as the presence of an AGN or other minor scale variations that exceed the scope of primary parameters such as the stellar mass or morphology. Conclusively, we propose that CL can serve as an embedding function for the development of larger models capable of integrating data from multiple datasets, thereby advancing the construction of more comprehensive foundation models for galaxies.

Improved source classification and performance analysis using Gaia DR3

The Discrete Source Classifier (DSC) provides probabilistic classification of sources in Gaia Data Release 3 (GDR3) using a Bayesian framework and a global prior. The DSC Combmod classifier in GDR3 achieved for the extragalactic classes (quasars and galaxies) a high completeness of 92%, but a low purity of 22% (all sky, all magnitudes) due to contamination from the far larger star class. However, these single metrics mask significant variation in performance with magnitude and sky position. Furthermore, a better combination of the individual classifiers that comprise Combmod is possible. Here we compute two-dimensional (2D) representations of the completeness and the purity as a function of Galactic latitude and source brightness, and also exclude the Magellanic Clouds where stellar contamination significantly reduces the purity. Reevaluated on a cleaner validation set and without introducing changes to the published GDR3 DSC probabilities themselves, we here achieve for Combmod average 2D completenesses of 92% and 95% and average 2D purities of 55% and 89% for the quasar and galaxy classes, respectively. Since the relative proportions of extragalactic objects to stars in Gaia is expected to vary significantly with brightness and latitude, we then introduce a new prior that is a continuous function of brightness and latitude, and compute new class probabilities from the GDR3 DSC component classifiers, Specmod and Allosmod. Contrary to expectations, this variable prior only improves the performance by a few percentage points, mostly at the faint end. Significant improvement, however, is obtained by a new additive combination of Specmod and Allosmod. This classifier, Combmod-α, achieves average 2D completenesses of 82% and 93% and average 2D purities of 79% and 93% for the quasar and galaxy classes, respectively, when using the global prior. Thus, we achieve a significant improvement in purity for a small loss of completeness. The improvement is most significant for faint quasars (G≥20) where the purity rises from 20% to 62%.

The fountain of the luminous infrared galaxy Zw049.057 as traced by its OH megamaser

High-resolution ($0 $10-35$\,pc ) e-MERLIN ($ cm) and ($0 $6.5$\,pc ) ALMA ($ 1.1$ mm) observations have been used to image OH (hydroxyl) and H$_2$CO (formaldehyde) megamaser emission, and HCN $3 2$ emission toward the nuclear ($&lt;100$ pc) region of the luminous infrared galaxy Zw049.057. Zw049.057 hosts a compact obscured nucleus (CON), and thus represents a class of galaxies that are often associated with inflow and outflow motions. Formaldehyde megamaser emission has been detected toward the nuclear region, $ 30$ pc ($0 and traces a structure along the disk major axis. OH megamaser (OHM) emission has been detected along the minor axis of the disk, $ 30$\,pc ($0 from the nucleus, where it exhibits a velocity gradient with extrema of $-20$ $ northwest (NW) of the disk. HCN $3 2$ emission reveals extended emission, along the disk minor axis out to $ 60$\,pc ($0 Analysis of the minor axis HCN emission reveals high-velocity features extending out to $) and collimated outflow that is enveloped by a wide-angle and slow ($ $) outflow that is traced by the OHM emission. Analysis of the outflow kinematics suggests that the slow wide-angle outflow will not reach escape velocity and will instead fall back to the galaxy disk, evolving as a so-called fountain flow, while the fast collimated outflow traced by HCN emission will likely escape the nuclear region. We suggest that the absence of OHM emission in the nuclear region is due to high densities there. Even though OHMs associated with outflows are an exception to conventional OHM emission, we expect them to be common in CON sources that host both OHM and H$_2$CO megamasers.

ALMA-ALPINE [CII] survey: The sub-kpc morphology of three main sequence galaxy systems at z ∼ 4.5 revealed by ALMA

Context. Going from a redshift of 6 down to nearly 4, galaxies grow rapidly from low-mass galaxies towards the more mature types of massive galaxies seen at cosmic noon. Growth via gas accretion and mergers undoubtedly shape this evolution, however, there is considerable uncertainty at present over the contribution of each of these processes to the overall evolution of galaxies. Furthermore, previous characterisations of the morphology of galaxies in the molecular gas phase have been limited by the coarse resolution of earlier observations. Aims. In this work, we utilise new high-resolution ALMA [CII] observations to analyse three main sequence (MS) galaxy systems at a redshift of z ∼ 4.5 and at resolutions of up to 0.15″. This approach enables us to investigate the morphology and kinematics on a kpc scale and understand the processes at play as well as the classifications of galaxies at high resolution. Thanks to this unique window, we are able to gain insights into the molecular gas of MS galaxies undergoing mass assembly in the early Universe. Methods. We used intensity and velocity maps, position-velocity diagrams, and radial profiles of [CII] in combination with dust continuum maps to analyse the morphology and kinematics of the three systems. Results. In general, we find that the high-resolution ALMA data reveal more complex morpho-kinematic properties. For one galaxy in our sample, we identified interaction-induced clumps, demonstrating the profound effect that mergers have on the molecular gas in galaxies, which is consistent with what has been suggested by recent simulations. One galaxy that was previously classified as dispersion-dominated turned out to show two bright [CII] emission regions, which could either be classified as merging galaxies or massive star-forming regions within the galaxy itself. The high-resolution data for the other dispersion dominated object also revealed clumps of [CII] that had not been identified previously. Within the sample, we might also detect star-formation powered outflows (or outflows from active galactic nuclei) that appear to be fuelling diffuse gas regions and enriching the circumgalactic medium. The new high-resolution ALMA data we present in this paper reveal that the galaxies in our sample are much more complex than they previously appeared in the low-resolution ALPINE data. In particular, we find evidence of merger induced clumps in the galaxy DC8187, along with signs of merging components for the other two objects. This may be evidence that the number of mergers at high redshift are significantly underestimated at present.

Galaxies in the zone of avoidance: Misclassifications using machine learning tools

Context. Automated methods for classifying extragalactic objects in large surveys offer significant advantages compared to manual approaches in terms of efficiency and consistency. However, the existence of the Galactic disk raises additional concerns. These regions are known for high levels of interstellar extinction, star crowding, and limited data sets and studies. Aims. In this study, we explore the identification and classification of galaxies in the zone of avoidance (ZoA). In particular, we compare our results in the near-infrared (NIR) with X-ray data. Methods. We analyzed the appearance of objects in the Galactic disk classified as galaxies using a published machine-learning (ML) algorithm and make a comparison with the visually confirmed galaxies from the VVV NIRGC catalog. Results. Our analysis, which includes the visual inspection of all sources cataloged as galaxies throughout the Galactic disk using ML techniques reveals significant differences. Only four galaxies were found in both the NIR and X-ray data sets. Several specific regions of interest within the ZoA exhibit a high probability of being galaxies in X-ray data but closely resemble extended Galactic objects. Our results indicate the difficulty in using ML methods for galaxy classification in the ZoA, which is mainly due to the scarcity of information on galaxies behind the Galactic plane in the training set. They also highlight the importance of considering specific factors that are present to improve the reliability and accuracy of future studies in this challenging region.

Two Distinct Classes of Quiescent Galaxies at Cosmic Noon Revealed by JWST PRIMER and UNCOVER

We present a measurement of the low-mass quiescent size–mass relation at cosmic noon (1 < z < 3) from the JWST PRIMER and UNCOVER treasury surveys, which highlights two distinct classes of quiescent galaxies. While the massive population is well studied at these redshifts, the low-mass end has been previously underexplored due to a lack of observing facilities with sufficient sensitivity and spatial resolution. We select a conservative sample of low-mass quiescent galaxy candidates using rest-frame UVJ colors and specific star formation rate criteria and measure galaxy morphology in both rest-frame UV/optical wavelengths (F150W) and rest-frame near-infrared (F444W). We confirm an unambiguous flattening of the low-mass quiescent size–mass relation, which results from the separation of the quiescent galaxy sample into two distinct populations at log(M⋆/M⊙)∼10.3 : low-mass quiescent galaxies that are notably younger and have disky structures, and massive galaxies consistent with spheroidal morphologies and older median stellar ages. These separate populations imply mass quenching dominates at the massive end while other mechanisms, such as environmental or feedback-driven quenching, form the low-mass end. This stellar mass-dependent slope of the quiescent size–mass relation could also indicate a shift from size growth due to star formation (low masses) to growth via mergers (massive galaxies). The transition mass between these two populations also corresponds with other dramatic changes and characteristic masses in several galaxy evolution scaling relations (e.g., star formation efficiency, dust obscuration, and stellar-to-halo mass ratios), further highlighting the stark dichotomy between low-mass and massive galaxy formation.

Improved Galaxy Morphology Classification with Convolutional Neural Networks

The increased volume of images and galaxies surveyed by recent and upcoming projects consolidates the need for accurate and scalable automated AI-driven classification methods. This paper proposes a new algorithm based on a custom neural network architecture for classifying galaxies from deep space surveys. The convolutional neural network (CNN) presented is trained using 10,000 galaxy images obtained from the Galaxy Zoo 2 dataset. It is designed to categorize galaxies into five distinct classes: completely round smooth, in-between smooth (falling between completely round and cigar-shaped), cigar-shaped smooth, edge-on, and spiral. The performance of the proposed CNN is assessed using a set of metrics such as accuracy, precision, recall, F1 score, and area under the curve. We compare our solution with well-known architectures like ResNet-50, DenseNet, EfficientNet, Inception, MobileNet, and one proposed model for galaxy classification found in the recent literature. The results show an accuracy rate of 96.83%, outperforming existing algorithms.

Exploiting Convolutional Neural Networks for Galaxy Classification

Galaxy classification is an important work in the field of astronomy. The efficiency of manually solving this problem is too low, and deep learning provides many methods for image classification. Using deep learning methods to automatically classify galaxies can greatly improve the efficiency of this task. Therefore, this paper uses mature Convolutional Neural Networks (CNN) models and self-designed simple CNN models, and compares their performance on the Galaxy10 DECals Dataset. In the experiment, the Inception network performed best, with an accuracy rate of 0.63 on the test set, but the training time was as long as about 2 hours. The simple CNN model ranked second, with an accuracy rate of 0.54, and the training time was only 20 minutes. The accuracy of other models is about 0.5, and the training time also takes 2 hours. The result performance of the model is worse than expected, which may be due to the less data used. The result of Inception is much ahead, and its solution to the problem provides a solution for the task of dealing with small data sets.

Mitigating bias in deep learning: training unbiased models on biased data for the morphological classification of galaxies

ABSTRACT Galaxy morphologies and their relation with physical properties have been a relevant subject of study in the past. Most galaxy morphology catalogues have been labelled by human annotators or by machine learning models trained on human-labelled data. Human-generated labels have been shown to contain biases in terms of the observational properties of the data, such as image resolution. These biases are independent of the annotators, that is, are present even in catalogues labelled by experts. In this work, we demonstrate that training deep learning models on biased galaxy data produces biased models, meaning that the biases in the training data are transferred to the predictions of the new models. We also propose a method to train deep learning models that considers this inherent labelling bias, to obtain a de-biased model even when training on biased data. We show that models trained using our deep de-biasing method are capable of reducing the bias of human-labelled data sets.