SLIDE 1
@kevinschawinski
ETH black hole group
Grüp Bœgg Negar Politecnic da Zürig
Artificial Intelligence meets Data Driven Astrophysics
Kevin Schawinski
Institute for Particle Physics and Astrophysics ETH Zurich
Artificial Intelligence meets Data Driven Astrophysics Kevin - - PowerPoint PPT Presentation
Artificial Intelligence meets Data Driven Astrophysics Kevin - - PowerPoint PPT Presentation
Artificial Intelligence meets Data Driven Astrophysics Kevin Schawinski Institute for Particle Physics and Astrophysics ETH Zurich ETH black hole group @kevinschawinski Grp Bgg Negar Politecnic da Zrig how can machine learning/
SLIDE 2
SLIDE 3
SLIDE 4
SLIDE 5
how can machine learning/ artificial intelligence help us understand the universe?
SLIDE 6
Large Synoptic Sky Telescope
early 2020s, 20 TB/night
SLIDE 7
Square Kilometer Array
mid 2020s, 160 TB/second
SLIDE 8
how can machine learning/ artificial intelligence help us understand the universe? do the simplest thing that works!
SLIDE 9
SLIDE 10
GalaxyGAN: de-noising and feature reconstruction PSFGAN: point source subtraction Generative models: data-driven exploration
SLIDE 11
GalaxyGAN: de-noising and feature reconstruction PSFGAN: point source subtraction Generative models: data-driven exploration
SLIDE 12
SLIDE 13
SLIDE 14
SLIDE 15
SLIDE 16
R ~ λ/D
resolving power of a telescope
λ - wavelength at which you observe D - diameter of your telescope aperture
SLIDE 17
Original Image Degraded Image Artificial Degrading
Data Prep. Training of GAN
Recovered Image Generator Original Image Discriminator (Original Image, Degraded Image) or (Recovered Image, Degraded Image)
generative adversarial network for overcoming limitations in astrophysical images
Schawinski+17
SLIDE 18
- riginal
degraded GAN recovered deconvolved PSF=2.5”, 10σ
Schawinski+17
SLIDE 19
low-resolution radio data NVSS survey high resolution ground truth FIRST survey GAN prediction
Semester project Nina Glaser Glaser et al. in prep.
SLIDE 20
GalaxyGAN: de-noising and feature reconstruction PSFGAN: point source subtraction Generative models: data-driven exploration
SLIDE 21
GalaxyGAN: de-noising and feature reconstruction PSFGAN: point source subtraction Generative models: data-driven exploration
SLIDE 22
SLIDE 23
SLIDE 24
SLIDE 25
Training Architecture Discriminator Generator Preprocessing Recovered Original Original + AGN
PSFGAN, Stark+ in press
Dominik Stark
SLIDE 26
PSFGAN, Stark+ in press
GAN performance vesus SOTA parametric fitting tool GALFIT
SLIDE 27
Less sensitive to PSF changes
PSFGAN, Stark+ in press
SLIDE 28
PSFGAN, Stark+ submitted
Less sensitive to training set
SLIDE 29
GalaxyGAN: de-noising and feature reconstruction PSFGAN: point source subtraction Generative models: data-driven exploration
SLIDE 30
GalaxyGAN: de-noising and feature reconstruction PSFGAN: point source subtraction Generative models: data-driven exploration
SLIDE 31
SLIDE 32
Dennis Turp
Encoder-decoder structure with a domain adversarial aspect
Fader network, Lample+17
SLIDE 33
Encoder-decoder structure with a domain adversarial aspect
SLIDE 34
SLIDE 35
Hypothesis generation
SLIDE 36
where we are where we want to get to
SLIDE 37
where we are where we want to get to
SLIDE 38
changing SSFR in latent space changing bulge-to-disk in latent space
SLIDE 39
go to space.ml to try out our projects!