Oral Presentation
Where's Swimmy?: Subaru wide-field machine-learning anomaly survey
Presenter: Rhythm Shimakawa (NAOJ)
As astronomical data rapidly grow in size and complexity these days, it is necessary to manipulate and make effective use of such big data-sets. A machine-learning approach is a novel technique that makes data mining quite effective, and its application to anomaly detection enables a systematic search of exceptionally rare sources in our universe.
This talk introduces our deep-learning anomaly survey based on multi-colored images taken from Subaru Strategic Program with Hyper-Suprime Cam (HSC-SSP S20A=>PDR3), called "SWIMMY" (Subaru WIde-field Machine-learning anoMaly surveY). In this project, we first establish a neural network through denoising autoencoder which reconstructs galaxy images in detail but without uncommon features in them. Then, scrutinizing residuals between original and reconstructed images allows us to extract abnormal features contained only in rare objects from the entire sample.
Initial results show that our early-stage model successfully detects anomalous imaging features of such as quasars and extreme emission-line galaxies. An application of this model to the whole HSC-SSP data will achieve a deep, comprehensive outlier search for millions of astronomical sources in the universe.
asiaa.sinica.edu.tw