You are here

Princeton University Data Science / COMPASS Seminar

Learning to See Like HSC & LSST

Abstract: LSST and HSC produce large and deep imaging data, for which detection and blending are challenging. For reliable cosmological interpretation of their data, we need to produce mock catalogs that capture all relevant selection effects, which usually requires expensive image simulations. I will present an emulator, which predicts a realistic detection catalog for a specific survey based on an input source catalog (from e.g. a N-body simulation) without the need for image simulations. The emulator is based on a novel deep clustering method, which uses a neural network to act as kernel function for the mean-shift clustering algorithm. By training the network in supervised fashion on actual detection catalogs, it learns to approximate the underlying density distribution. In essence: it learns to see the sky like a particular instrument would.  Every week we discuss data science methods and applications from papers, reviews, software releases, etc. We also have demos for useful/fancy methods by locals and visitors through the COMPASS program. The setting is informal. Material should be presented directly from the source or on the white board, demos should be hands-on. We collect links to documents, source code, tutorials, etc. for later perusal in this github repo. If you would like to present at the data science seminar, please contact the organizers Peter Melchior (melchior[at], Adrian Price-Whelan (adrn[at], Christina Kreisch (ckreisch[at], Lachlan Lancaster (lachlanl[at]


Peter Melchior

Speaker Affiliation

Princeton University


Natural Sciences, Natural Sciences

Additional Information

Date & Time
April 15, 2019 | 4:005:00pm


Peyton Hall, Room 140