Self Organizing Neural Networks: From Galaxy Evolution to Cosmology

Shoubaneh Hemmati (IPAC)
Friday, February 2, 2018 - 9:15am

In this talk I utilize Self Organizing Maps (SOMs), a class of unsupervised neural networks, for dealing with specific challenges in astrophysics from meeting the cosmological survey requirements to measuring galaxy properties. To infer cosmological parameters not limited by systematic errors, WFIRST and other Stage IV Dark energy experiments (e.g. LSST, Euclid) need very accurate redshift measurements. This accuracy can only be met using spectroscopic subsamples to calibrate the full sample. Using the galaxies from the CANDELS survey I build the LSST+WFIRST lensing analog sample of ~36k objects and train the LSST+WFIRST SOM. We find 26% of the WFIRST lensing sample consists of sources fainter than the Euclid depth in the optical, 91% of which live in color cells already occupied by brighter galaxies. 4% of SOM cells are however only occupied by faint galaxies for which we recommend extra spectroscopy enabling the comprehensive calibration for the LSST+WFIRST lensing sample. I continue by demonstrating how SOMs can be used to visualize and optimize libraries of stellar evolution synthesis models. SOMs are able to provide not only a much quicker measurement of the galaxy properties but also provide better constraints on the uncertainties compared to traditional SED fitting. This is essential for processing the massive datasets expected from next generation galaxy surveys.

Talk Type: