Artificial Intelligence meets Data Driven Astrophysics Kevin Schawinski Institute for Particle Physics and Astrophysics ETH Zurich ETH black hole group @kevinschawinski Grüp Bœgg Negar Politecnic da Zürig
how can machine learning/ artificial intelligence help us understand the universe?
Large Synoptic Sky Telescope early 2020s, 20 TB/night
Square Kilometer Array mid 2020s, 160 TB/second
how can machine learning/ artificial intelligence help us understand the universe? do the simplest thing that works!
GalaxyGAN: de-noising and feature reconstruction PSFGAN: point source subtraction Generative models: data-driven exploration
GalaxyGAN: de-noising and feature reconstruction PSFGAN: point source subtraction Generative models: data-driven exploration
resolving power of a telescope R ~ λ /D λ - wavelength at which you observe D - diameter of your telescope aperture
generative adversarial network for overcoming limitations in astrophysical images Data Prep. Training of GAN Original Image Original Image (Original Image, Degraded Image) or (Recovered Image, Degraded Image) Discriminator Artificial Degrading Degraded Image Generator Recovered Image Schawinski+17
original degraded GAN recovered deconvolved PSF=2.5”, 10 σ Schawinski+17
low-resolution radio data high resolution ground truth GAN prediction NVSS survey FIRST survey Semester project Nina Glaser Glaser et al. in prep.
GalaxyGAN: de-noising and feature reconstruction PSFGAN: point source subtraction Generative models: data-driven exploration
GalaxyGAN: de-noising and feature reconstruction PSFGAN: point source subtraction Generative models: data-driven exploration
Training Architecture Original Discriminator Preprocessing Original + AGN Recovered Generator Dominik Stark PSFGAN, Stark+ in press
GAN performance vesus SOTA parametric fitting tool GALFIT PSFGAN, Stark+ in press
Less sensitive to PSF changes PSFGAN, Stark+ in press
Less sensitive to training set PSFGAN, Stark+ submitted
GalaxyGAN: de-noising and feature reconstruction PSFGAN: point source subtraction Generative models: data-driven exploration
GalaxyGAN: de-noising and feature reconstruction PSFGAN: point source subtraction Generative models: data-driven exploration
Encoder-decoder structure with a domain adversarial aspect Dennis Turp Fader network, Lample+17
Encoder-decoder structure with a domain adversarial aspect
Hypothesis generation
where we want to where we are get to
where we want to where we are get to
changing SSFR changing bulge-to-disk in latent space in latent space
go to space.ml to try out our projects!
Recommend
More recommend
