Euclid Consortium Annual Meeting 2016, Lisbon

Transcript

1. Galaxy Zoo Express: What the combined power of human and machine can do for Euclid Melanie Beck University of Minnesota 2 June, 2016 beck@astro.umn.edu
2. Classifying morphologies → Visual: Too slow → Standard machines: Not as accurate → Deep learning: massive training samples, hard to interpret
3. Classifying morphologies → Visual: Too slow → Standard machines: Not as accurate → Deep learning: massive training samples, hard to interpret A machine is just an algorithm. (GIGO) Both humans and machines can be trained. Fewer classes are easier to predict than many.
4. HVPU adapted from SpaceWarps, Marshall+2016
5. HVPU adapted from SpaceWarps, Marshall+2016
6. HVPU adapted from SpaceWarps, Marshall+2016
7. HVPU adapted from SpaceWarps, Marshall+2016
8. Simulations on Galaxy Zoo 2 285,962 SDSS galaxies 83,943 volunteers 14 months project completion Willett+2013 ‘Smooth’ ‘Not Smooth’
10. Simple Machines 1. Random Forest (scikit-learn) 2. Fast to train 3. Require modestly-sized training samples 4. Feature vector: (C, A, M20, G, e)
11. Preliminary performance “Days saved”, “human hours saved” HVPU Machine
12. Can we tackle the Euclid challenge? GZ currently classifies ~350 galaxies per day. Euclid could provide upwards of 107 – 108 galaxies with measurable morphology. Will take 700+ years to visually classify all galaxies. GZX reclassifed GZ2 in ~40 “GZ2 days” instead of 14 months, an order of magnitude improvement.
13. Can we tackle the Euclid challenge? GZ currently classifies ~350 galaxies per day. Euclid could provide upwards of 107 – 108 galaxies with measurable morphology. Will take 700+ years to visually classify all galaxies. GZX reclassifed GZ2 in ~40 “GZ2 days” instead of 14 months, an order of magnitude improvement. Yes...
14. What does GZX need? 1.Human & Machine classifiers 2.Features that correlate with morphology (already in pipeline) 3.JPEG color-composite images of galaxies • Monochrome imaging (VIS)? That's OK. (Lintott+2008) • Do other Euclideans need this as well? 4.Expert-classified sample • Train volunteers • Validate machines Can possibly use current expertly-classified samples (CANDELS, Nair&Abraham, etc.) or new GZ samples (GZH & GZC)
15. In progress 1.Apply to Illustris simulated galaxies: a.We have Galaxy Zoo classifications for ~20K (running more...) b.Lotz measured CAS, G-M on these 2.Test feature manipulation 3.Test other morphology questions 4.Run live version
16. Closing remarks GZ currently GZX quickly provides classifications at the 80-85% accuracy level in an order of magnitude less time than visual classifications alone. We can isolate which galaxies are difficult to classify and direct them to the most sensitive classifiers (humans). GZX is built to be modular and will soon incorporate any machine (standard algorithms or deep learning instances) as well as ensembles of machines. GZX can accommodate any set of feature vectors that correlate with morphology. (Systematics in features can be handled)
17. Space Warps Analysis Pipeline Marshall+2016