The Subaru Seminar is
usually held in Room 104 of the Hilo Base Facility, adjacent
to the main lobby. Everyone is welcome to attend. If you are
interested in giving a seminar, please contact Subaru seminar organizers
(Tae-Soo Pyo, Sherry Yeh, Nagayoshi Ohashi)
by email : sseminar_at_subaru.naoj.org (please change"_at" to @).
December 09, Wednesday, 03:00 pm in 104
" A Brief Introduction of Deep Neural Network and its Application to Astronomy "
(NTT Communication Science Laboratories)
I present our work on applying the Deep Neural Network (DNN) to an astronomy research project. DNN is the hottest topic in machine learning researches, due to its record-breaking performance in several machine learning problems. After introducing DNN briefly, we present our application of DNN to help astronomy researchers to identify "supernova" and "bogus (non-star)" image patches taken by the SUBARU telescope. Our DNN model is trained to classify an image patch into a supernova class or a bogus class, observing a collection of measurements extracted from image patches shot by the SUBARU HSC. The classification performance of DNN is superior to existing machine learning techniques, even if the number of observation is not so much large. An automatic filtering of image by machine learning techniques will reduce efforts of astronomy researchers very much in choosing supernova candidates to be tested with spectrum analyzers.
" Machine Learning Selection for Supernovae from Subaru/HSC data "
(Institute of Statistics and Math)
Since 2014, a large survey project using Subaru/HSC(300 nights for 5 years) has started to study the cosmology and to search for optical transients. The problem in this survey is to find real transients from candidates including a large amount of bogus objects in difference images. We present the process of selection of supernovae using machine learning techniques.
I present the boosting method, and Ishiguro-san shows the Deep Neural Network.
Seminars are also held at JAC,