A Typical User's Ground-Level Perspective on Machine Learning in Astronomy

Slides from the AAS Splinter talk I gave on machine learning for normal astronomers.

The full talk video was also posted here:
https://www.youtube.com/watch?v=jaB5jghcu20

Full Abstract:
We have entered an era in observational astronomy in which sky surveys routinely release massive datasets. While this wealth of data is critical for determining rates of rare phenomena (e.g. transiting exoplanets or tidal disruption events), it also enables a new kind of data-driven astrophysics (e.g. "hidden" correlations in our data that point towards new or challenging undetandings of physics). Machine learning is simply one tool available to us to discover these new trends or make predictions from our growing volume of data. However, machine learning alone cannot make astrophysical discoveries, and astronomers are still required to interpret astrophysical meaning from our data. Here I will discuss some uses of machine learning in analyzing data from the Kepler and Gaia missions, and attempt to highlight some of the opportunities and limitations in its use.

Transcript

1. A Typical User's Ground-Level Perspective on Machine Learning in Astronomy James R. A. Davenport jradavenport UNIVERSITY OF WASHINGTON !1
2. A Typical User's Ground-Level Perspective on Machine Learning in Astronomy jradavenport !2 Number of ADS abstracts including “machine learning”
3. jradavenport !3 Machine Learning is Boring*
4. jradavenport !4 *normal, lots of people doing it. And this is a good thing! (i.e. don’t be scared!) Machine Learning is Boring*
5. jradavenport !5 (i.e. can run most algorithms on your laptop) Yet sample size is big enough that interesting/rare things can be found Most of our work isn’t “big data”
6. jradavenport !6 Our data is becoming better suited for ML • big datasets (Mario’s talk), especially Gaia! • easier than ever to get data (Vizier/Xmatch, ADS, journals, Github, Zenodo…) • value-added datasets for surveys (e.g. stellar parameters from SDSS)
7. jradavenport !7 ML is easier than ever to use • robust, open source libraries available • many programming languages • many domain (astro) experts & workshops available
8. jradavenport !8
9. jradavenport !9 also astroML! 
 (see Brigitta’s talk later)
10. jradavenport !10 Problems that ML is good for:
11. jradavenport !11 Each algorithm has specific use cases (sometimes: just try them all!) Clustering Classification
12. jradavenport !12 Goal: demonstrate 3 real problems that ML can be used for
13. jradavenport !13 Example 1: Clustering starspot evolution tracks Kepler 17
14. jradavenport !14 Kepler 17 Example 1: Clustering starspot evolution tracks
15. jradavenport !15 Kepler 17 Example 1: Clustering starspot evolution tracks
16. jradavenport !16 Question: Which tracks are starspots, how do they emerge/decay? Kepler 17 Example 1: Clustering starspot evolution tracks
17. jradavenport !17 Question: Which tracks are starspots, how do they emerge/decay? Manual clustering? Kepler 17 Example 1: Clustering starspot evolution tracks
18. jradavenport !18 Question: Which tracks are starspots, how do they emerge/decay? Manual clustering? Kepler 17 Example 1: Clustering starspot evolution tracks Aside: “training data” is super 
 important for many problems!
19. Example 1: Clustering starspot evolution tracks jradavenport !19 DBSCAN Gaussian Mixture
20. Example 1: Clustering starspot evolution tracks jradavenport !20 DBSCAN Gaussian Mixture need to predefine Nclusters?
21. jradavenport !21 DBSCAN: Density-based spatial clustering of applications with noise Kepler 17 Example 1: Clustering starspot evolution tracks
22. jradavenport !22 DBSCAN: Density-based spatial clustering of applications with noise One starspot’s emergence/decay! Kepler 17 See more upcoming work by Kosuke Namekata Example 1: Clustering starspot evolution tracks
23. Time Flux jradavenport !23 Example 2: Modeling a complex stellar flare Flare! Question: Is there (quasi-) sinusoidal behavior in the flare decay? https://github.com/RileyWClarke/QPP-GP
24. Time Flux jradavenport !24 Flare! Question: Is there (quasi-) sinusoidal behavior in the flare decay? https://github.com/RileyWClarke/QPP-GP versus Example 2: Modeling a complex stellar flare
25. jradavenport !25 Time Flux Only study decay Residual Flux Time Could fit with a damped 
 harmonic oscillator Question: Is there (quasi-) sinusoidal behavior in the flare decay? Difficult to classify sinusoidal vs. stochastic,
 & strict vs quasi sinusoid https://github.com/RileyWClarke/QPP-GP Example 2: Modeling a complex stellar flare
26. jradavenport !26 Residual Flux https://github.com/RileyWClarke/QPP-GP Celerite Gaussian Process Example 2: Modeling a complex stellar flare
27. jradavenport !27 Residual Flux Gaussian Process https://github.com/RileyWClarke/QPP-GP Celerite Use an exponential + 
 simple-harmonic-oscillator kernel Example 2: Modeling a complex stellar flare
28. jradavenport !28 https://github.com/RileyWClarke/QPP-GP No Period Candidate Periods Objective search Robust uncertainties!
 (with MCMC) Gaussian Process Example 2: Modeling a complex stellar flare
29. jradavenport !29 Example 3: Modeling photometric metallicities From APOGEE [3.4] - [4.6] Observation: [Fe/H] gradient in stars https://github.com/jradavenport/ingot/ Gaia DR2 + (WISE + 2MASS) + APOGEE
30. jradavenport !30 From APOGEE [3.4] - [4.6] Observation: [Fe/H] gradient in stars We could build a complex 
 polynomial or spline model https://github.com/jradavenport/ingot/ Example 3: Modeling photometric metallicities
31. jradavenport !31 Observation: [Fe/H] gradient in stars From APOGEE [3.4] - [4.6] We could build a complex 
 polynomial or spline model Tedious, and difficult to add
 additional dimensions! https://github.com/jradavenport/ingot/ Example 3: Modeling photometric metallicities
32. jradavenport !32 Observation: [Fe/H] gradient in stars From APOGEE [3.4] - [4.6] We could build a complex 
 polynomial or spline model Tedious, and difficult to add
 additional dimensions! Or use a simple, flexible 
 ML model! https://github.com/jradavenport/ingot/ Example 3: Modeling photometric metallicities
33. jradavenport !33 KNearestNeighbors Xdata = (G-J, W1-W2)
 Ydata = [Fe/H] Data Fit https://github.com/jradavenport/ingot/ Example 3: Modeling photometric metallicities
34. jradavenport !34 KNearestNeighbors Xdata = (G-J, W1-W2)
 Ydata = [Fe/H] Data Fit https://github.com/jradavenport/ingot/ Example 3: Modeling photometric metallicities
35. jradavenport !35 KNearestNeighbors Result: a simple to use “surface”,
 no tweaking for shape/order, 
 extend to addtional dimensions easily 1 Million new stars with no spectra https://github.com/jradavenport/ingot/ Example 3: Modeling photometric metallicities
36. jradavenport !36 KNearestNeighbors Result: a simple to use “surface”,
 no tweaking for shape/order, 
 extend to addtional dimensions easily 1 Million new stars with no spectra https://github.com/jradavenport/ingot/ Example 3: Modeling photometric metallicities
37. !37 Conclusions Clustering Regression GP’s ML is easier and more “boring” than ever! jradavenport