Effectively using unsupervised machine learning in next generation astronomical surveys

Astronomers have typically set out to solve supervised machine learning problems by creating their own representations from scratch. We show that deep learning models trained to answer every Galaxy Zoo DECaLS question learn meaningful semantic representations of galaxies that are useful for new tasks on which the models were never trained. We exploit these representations to outperform several recent approaches at practical tasks crucial for investigating large galaxy samples. The first task is identifying galaxies of similar morphology to a query galaxy. Given a single galaxy assigned a free text tag by humans (e.g. ‘#diffuse’), we can find galaxies matching that tag for most tags. The second task is identifying the most interesting anomalies to a particular researcher. Our approach is 100 per cent accurate at identifying the most interesting 100 anomalies (as judged by Galaxy Zoo 2 volunteers). The third task is adapting a model to solve a new task using only a small number of newly-labelled galaxies. Models fine-tuned from our representation are better able to identify ring galaxies than models fine-tuned from terrestrial images (ImageNet) or trained from scratch. We solve each task with very few new labels; either one (for the similarity search) or several hundred (for anomaly detection or fine-tuning). This challenges the longstanding view that deep supervised methods require new large labelled datasets for practical use in astronomy. To help the community benefit from our pretrained models, we release our fine-tuning code zoobot. Zoobot is accessible to researchers with no prior experience in deep learning.

show abstract

“…umap is commonly used for visualising high-dimensional spaces in both computer science and astronomy (e.g. Clarke et al 2020;Reis et al 2021).…”

Section: Visualisingmentioning

confidence: 99%

Practical galaxy morphology tools from deep supervised representation learning

Walmsley

Scaife

Lintott

et al. 2022

Monthly Notices of the Royal Astronomical Society

View full text Add to dashboard Cite

show abstract

“…This indicates that our approach could be used to robustly filter out bad images at the pipeline level, in addition to identifying scientifically interesting anomalies in postprocessing. We built a custom visualization tool to interactively explore the UMAP space in more detail, based on a similar tool by Reis et al (2021); it can be accessed at https://weirdgalaxi.es. We used this tool to perform a search for scientifically interesting anomalies; the results of this search are described in Section 4.5.…”

Section: Characterization Of Anomalies With the Wgan Cae And Umapmentioning

confidence: 99%

Anomaly detection in Hyper Suprime-Cam galaxy images with generative adversarial networks

Storey-Fisher,

Huertas-Company,

Ramachandra

et al. 2021

Preprint

View full text Add to dashboard Cite

The problem of anomaly detection in astronomical surveys is becoming increasingly important as data sets grow in size. We present the results of an unsupervised anomaly detection method using a Wasserstein generative adversarial network (WGAN) on nearly one million optical galaxy images in the Hyper Suprime-Cam (HSC) survey. The WGAN learns to generate realistic HSC-like galaxies that follow the distribution of the data set; anomalous images are defined based on a poor reconstruction by the generator and outlying features learned by the discriminator. We find that the discriminator is more attuned to potentially interesting anomalies compared to the generator, so we use the discriminator-selected images to construct a high-anomaly sample of ∼13,000 objects. We propose a new approach to further characterize these anomalous images: we use a convolutional autoencoder (CAE) to reduce the dimensionality of the residual differences between the real and WGAN-reconstructed images and perform clustering on these. We report detected anomalies of interest including galaxy mergers, tidal features, and extreme star-forming galaxies. We perform follow-up spectroscopy of several of these objects, and present our findings on an unusual system which we find to most likely be a metal-poor dwarf galaxy with an extremely blue, higher-metallicity HII region. We have released a catalog with the WGAN anomaly scores; the code and catalog are available at https://github.com/kstoreyf/anomalies-GAN-HSC, and our interactive visualization tool for exploring the clustered data is at https://weirdgalaxi.es.

show abstract

“…The former are concerned with extracting patterns from a data set without direct guidance in the form of labeled data, while the latter focus on learning a function from labeled examples to carry out classification or regression. Examples of the former are anomaly detection (Protopapas et al 2006;Baron & Poznanski 2017;Giles & Walkowicz 2020), clustering (e.g., Pasquato & Chung 2019), dimensionality reduction (e.g., Reis et al 2018), and even integrated approaches including interactive visualization (Reis et al 2021). While we do not discuss supervised methods in the following (nor even unsupervised methods except for SOM), we point the interested reader to two relevant reviews: Ball & Brunner (2010) and the more recent Baron (2019).…”

Section: Introductionmentioning

confidence: 99%

Exploring X-ray variability with unsupervised machine learning

Pasquato

Marelli

Luca

et al. 2022

A&A

View full text Add to dashboard Cite

Context. XMM-Newton provides unprecedented insight into the X-ray Universe, recording variability information for hundreds of thousands of sources. Manually searching for interesting patterns in light curves is impractical, requiring an automated data-mining approach for the characterization of sources. Aims. Straightforward fitting of temporal models to light curves is not a sure way to identify them, especially with noisy data. We used unsupervised machine learning to distill a large data set of light-curve parameters, revealing its clustering structure in preparation for anomaly detection and subsequent searches for specific source behaviors (e.g., flares, eclipses). Methods. Self-organizing maps (SOMs) achieve dimensionality reduction and clustering within a single framework. They are a type of artificial neural network trained to approximate the data with a two-dimensional grid of discrete interconnected units, which can later be visualized on the plane. We trained our SOM on temporal-only parameters computed from 10 5 detections from the EXTraS (Exploring the X-ray Transient and variable Sky) catalog. Results. The resulting map reveals that the ≈ 2500 most variable sources are clustered based on temporal characteristics. We find distinctive regions of the SOM map associated with flares, eclipses, dips, linear light curves, and others. Each group contains sources that appear similar by eye. We single out a handful of interesting sources for further study. Conclusions. The condensed view of our dataset provided by SOMs allowed us to identify groups of similar sources, speeding up manual characterization by orders of magnitude. Our method also highlights problems with fitting simple temporal models to light curves and can be used to mitigate them to an extent. This will be crucial for fully exploiting the high data volume expected from upcoming X-ray surveys, and may also help with interpreting supervised classification models.

show abstract

Effectively using unsupervised machine learning in next generation astronomical surveys

Cited by 14 publications

References 43 publications

Practical galaxy morphology tools from deep supervised representation learning

Practical galaxy morphology tools from deep supervised representation learning

Anomaly detection in Hyper Suprime-Cam galaxy images with generative adversarial networks

Exploring X-ray variability with unsupervised machine learning

Contact Info

Product

Resources

About