machine learning in the era of multiwavelength astronomygopira.jp/sym2020/2-10-takeuchi.pdf ·...
TRANSCRIPT
![Page 1: Machine Learning in the Era of Multiwavelength Astronomygopira.jp/sym2020/2-10-Takeuchi.pdf · Machine Learning in the Era of Multiwavelength Astronomy Tsutomu T. TAKEUCHI. Suchetha](https://reader033.vdocuments.us/reader033/viewer/2022051807/6003f66893f6c563617421fe/html5/thumbnails/1.jpg)
Machine Learning in the Era of Multiwavelength Astronomy
Tsutomu T. TAKEUCHI
Suchetha COORAY, Kai T. KONO, Wen E. SHI, Hai-Xia MA
Division of Particle and Astrophysical Science, Nagoya UniversityInstitute of Statistical Mathematics
GOPIRA Symposium 2020, Online, 15 Sep., 2020
![Page 2: Machine Learning in the Era of Multiwavelength Astronomygopira.jp/sym2020/2-10-Takeuchi.pdf · Machine Learning in the Era of Multiwavelength Astronomy Tsutomu T. TAKEUCHI. Suchetha](https://reader033.vdocuments.us/reader033/viewer/2022051807/6003f66893f6c563617421fe/html5/thumbnails/2.jpg)
1. Introduction: Multiwavelength Astronomy
Hertzsprung-Russel diagram
https://en.wikipedia.org/wiki/Hertzsprung%E2%80%93Russell_diagram
1.1 Traditional method
The most well-known exampleof the color-magnitude diagram.Stellar evolution was quantifiedvastly through the H-R diagram.
![Page 3: Machine Learning in the Era of Multiwavelength Astronomygopira.jp/sym2020/2-10-Takeuchi.pdf · Machine Learning in the Era of Multiwavelength Astronomy Tsutomu T. TAKEUCHI. Suchetha](https://reader033.vdocuments.us/reader033/viewer/2022051807/6003f66893f6c563617421fe/html5/thumbnails/3.jpg)
The theoretical evolutionarytrack of stars has beendeveloped by a detailedcomparison with observedH-R diagrams.
The H-R diagram has beena tremendously importanttool in astronomy.
Chiosi et al (1992)
H-R diagram and stellar evolutionary track
![Page 4: Machine Learning in the Era of Multiwavelength Astronomygopira.jp/sym2020/2-10-Takeuchi.pdf · Machine Learning in the Era of Multiwavelength Astronomy Tsutomu T. TAKEUCHI. Suchetha](https://reader033.vdocuments.us/reader033/viewer/2022051807/6003f66893f6c563617421fe/html5/thumbnails/4.jpg)
If we plot galaxy luminosity(absolute magnitude) vs.color, a clear dichotomy isfound: the color bimodality.
Redder galaxies:red sequence
Bluer galaxies:blue cloud
Boundary: green valley
Blanton (2006)
Absolute magnitude Mg
Col
or(u
-g)
Col
or (u
-g)
Color-magnitude relation of galaxies
![Page 5: Machine Learning in the Era of Multiwavelength Astronomygopira.jp/sym2020/2-10-Takeuchi.pdf · Machine Learning in the Era of Multiwavelength Astronomy Tsutomu T. TAKEUCHI. Suchetha](https://reader033.vdocuments.us/reader033/viewer/2022051807/6003f66893f6c563617421fe/html5/thumbnails/5.jpg)
Bruzual & Charlot (1993)
Flux
(no
rmal
ized
)
Wavelength [Å]
Star formation history (SFH)is one of the key factors ofgalaxy evolution.
SFH is directly reflected to thespectral luminosity of galaxies.
Galaxy evolution in multiwavelength
![Page 6: Machine Learning in the Era of Multiwavelength Astronomygopira.jp/sym2020/2-10-Takeuchi.pdf · Machine Learning in the Era of Multiwavelength Astronomy Tsutomu T. TAKEUCHI. Suchetha](https://reader033.vdocuments.us/reader033/viewer/2022051807/6003f66893f6c563617421fe/html5/thumbnails/6.jpg)
Flux
(no
rmal
ized
)
Wavelength [Å]
Star formation history (SFH)is one of the key factors ofgalaxy evolution.
SFH is directly reflected to thespectral luminosity of galaxies.
Galaxy evolution related to theSFH will be well represented inthe multiwavelength (band)luminosity space.
Galaxy evolution in multiwavelength
![Page 7: Machine Learning in the Era of Multiwavelength Astronomygopira.jp/sym2020/2-10-Takeuchi.pdf · Machine Learning in the Era of Multiwavelength Astronomy Tsutomu T. TAKEUCHI. Suchetha](https://reader033.vdocuments.us/reader033/viewer/2022051807/6003f66893f6c563617421fe/html5/thumbnails/7.jpg)
Colors are basically ratios of two luminosities.
⇒ Selection effect is always too entangled and messy.
⇒ Completeness test is almost impossible in a simple way.
1.2 Problem in the Color-Based Methods
![Page 8: Machine Learning in the Era of Multiwavelength Astronomygopira.jp/sym2020/2-10-Takeuchi.pdf · Machine Learning in the Era of Multiwavelength Astronomy Tsutomu T. TAKEUCHI. Suchetha](https://reader033.vdocuments.us/reader033/viewer/2022051807/6003f66893f6c563617421fe/html5/thumbnails/8.jpg)
Colors are basically ratios of two luminosities.
⇒ Selection effect is always too entangled and messy.
⇒ Completeness test is almost impossible in a simple way.
Suggestion: forget about colors!
Instead, we can simply use the distribution of galaxies ina multidimensional luminosity (absolute magnitude)space.
1.2 Problem in the Color-Based Methods
![Page 9: Machine Learning in the Era of Multiwavelength Astronomygopira.jp/sym2020/2-10-Takeuchi.pdf · Machine Learning in the Era of Multiwavelength Astronomy Tsutomu T. TAKEUCHI. Suchetha](https://reader033.vdocuments.us/reader033/viewer/2022051807/6003f66893f6c563617421fe/html5/thumbnails/9.jpg)
Since we have a bimodality in color-color space, wemust have an equivalent peaks in the multidimensionalluminosity space. Color-color plots only show reducedinformation.
From colors to multiwavelength luminosities
Color
Col
or
LuminosityL
umin
osity
![Page 10: Machine Learning in the Era of Multiwavelength Astronomygopira.jp/sym2020/2-10-Takeuchi.pdf · Machine Learning in the Era of Multiwavelength Astronomy Tsutomu T. TAKEUCHI. Suchetha](https://reader033.vdocuments.us/reader033/viewer/2022051807/6003f66893f6c563617421fe/html5/thumbnails/10.jpg)
2. What is the Galaxy Evolution?
Beatrice Tinsley
Galaxies evolve.
In those days
In galactic astronomy of 70-80’s, themeaning of galaxy evolution seemed tobe unambiguous.
Galaxy evolution= evolution of stellar population and chemical composition
![Page 11: Machine Learning in the Era of Multiwavelength Astronomygopira.jp/sym2020/2-10-Takeuchi.pdf · Machine Learning in the Era of Multiwavelength Astronomy Tsutomu T. TAKEUCHI. Suchetha](https://reader033.vdocuments.us/reader033/viewer/2022051807/6003f66893f6c563617421fe/html5/thumbnails/11.jpg)
Galaxies evolve in various aspects:
Today
![Page 12: Machine Learning in the Era of Multiwavelength Astronomygopira.jp/sym2020/2-10-Takeuchi.pdf · Machine Learning in the Era of Multiwavelength Astronomy Tsutomu T. TAKEUCHI. Suchetha](https://reader033.vdocuments.us/reader033/viewer/2022051807/6003f66893f6c563617421fe/html5/thumbnails/12.jpg)
x = x(T|T > t)
Today
Galaxies evolve in various aspects:
![Page 13: Machine Learning in the Era of Multiwavelength Astronomygopira.jp/sym2020/2-10-Takeuchi.pdf · Machine Learning in the Era of Multiwavelength Astronomy Tsutomu T. TAKEUCHI. Suchetha](https://reader033.vdocuments.us/reader033/viewer/2022051807/6003f66893f6c563617421fe/html5/thumbnails/13.jpg)
This is the formal and ultimate goal of the studies ongalaxy evolution, but clearly it is a substantiallycomplicated problem. It is time to define the evolution ofgalaxies with more objective point of view.
Today
Galaxies evolve in various aspects:
![Page 14: Machine Learning in the Era of Multiwavelength Astronomygopira.jp/sym2020/2-10-Takeuchi.pdf · Machine Learning in the Era of Multiwavelength Astronomy Tsutomu T. TAKEUCHI. Suchetha](https://reader033.vdocuments.us/reader033/viewer/2022051807/6003f66893f6c563617421fe/html5/thumbnails/14.jpg)
Historically, in 80’s, astronomers introduced a method ofclassical multivariate analysis such as PCA to find and unifyvarious scaling relations (e.g., Djorgovski 1992).
Djorgovski (1992)
However, since classical PCA-type analysis could only find alinear structure in the featurespace, the idea worked only to alimited problems and have beenonce forgotten.
3. Galaxy Manifold3.1 Galaxy manifold in the early days
![Page 15: Machine Learning in the Era of Multiwavelength Astronomygopira.jp/sym2020/2-10-Takeuchi.pdf · Machine Learning in the Era of Multiwavelength Astronomy Tsutomu T. TAKEUCHI. Suchetha](https://reader033.vdocuments.us/reader033/viewer/2022051807/6003f66893f6c563617421fe/html5/thumbnails/15.jpg)
Chilingarian et al. (2012)
Some preceding studies have suggested the existence of asmooth relation of galaxies in the 3D color–color–magnitudespace smoothly continuing from the blue cloud to the redsequence (e.g. Chilingarian et al. 2012).
⇒ general idea of a low dimensional submanifold existingin a higher dimensional feature space: revival of the galaxymanifold!
3.2 Galaxy manifold: reboot
![Page 16: Machine Learning in the Era of Multiwavelength Astronomygopira.jp/sym2020/2-10-Takeuchi.pdf · Machine Learning in the Era of Multiwavelength Astronomy Tsutomu T. TAKEUCHI. Suchetha](https://reader033.vdocuments.us/reader033/viewer/2022051807/6003f66893f6c563617421fe/html5/thumbnails/16.jpg)
• Reference Catalog of galaxy Spectral EnergyDistributions (RCSED) (Chilingarian et al. 2016)
• Catalog of galaxies produced as join between GALEX,SDSS, and UKIDSS catalogs, and processed with state-of-the-art spectral analysis methods
• Covers approximately 25% of the sky and contains k-corrected ultraviolet-to-near-infrared photometry (11bands of FUV, NUV, u, g, r, i, z, Y, J, H, K) of some 1million galaxies, as well as some of their physicalproperties
http://rcsed.sai.msu.ru
Galaxy manifold from RCSED
![Page 17: Machine Learning in the Era of Multiwavelength Astronomygopira.jp/sym2020/2-10-Takeuchi.pdf · Machine Learning in the Era of Multiwavelength Astronomy Tsutomu T. TAKEUCHI. Suchetha](https://reader033.vdocuments.us/reader033/viewer/2022051807/6003f66893f6c563617421fe/html5/thumbnails/17.jpg)
We want to characterize and parametrize this nonlinear manifold,and hopefully interpret physically!
Cooray et al. (2020)
Galaxy manifold from RCSED
![Page 18: Machine Learning in the Era of Multiwavelength Astronomygopira.jp/sym2020/2-10-Takeuchi.pdf · Machine Learning in the Era of Multiwavelength Astronomy Tsutomu T. TAKEUCHI. Suchetha](https://reader033.vdocuments.us/reader033/viewer/2022051807/6003f66893f6c563617421fe/html5/thumbnails/18.jpg)
Manifold Learning (or non-linear dimensionality reduction)embeds data that originally lies in a high dimensional space ina lower dimensional space, while preserving characteristicproperties.A manifold is a topological space that locally resemblesEuclidean space near each point.
3.3 Galaxy manifold from manifold learning
![Page 19: Machine Learning in the Era of Multiwavelength Astronomygopira.jp/sym2020/2-10-Takeuchi.pdf · Machine Learning in the Era of Multiwavelength Astronomy Tsutomu T. TAKEUCHI. Suchetha](https://reader033.vdocuments.us/reader033/viewer/2022051807/6003f66893f6c563617421fe/html5/thumbnails/19.jpg)
IsomapTenenbaum et al. (2000)
Neighboring points: input-space distance
Distant points: a sequence of “short hops” betweenneighboring points
Method: Finding shortest paths in a graph with edgesconnecting neighboring data points
N.B. Unlike the geodesic distance,the Euclidean distance cannotreflect the geometric structure ofthe data points.
![Page 20: Machine Learning in the Era of Multiwavelength Astronomygopira.jp/sym2020/2-10-Takeuchi.pdf · Machine Learning in the Era of Multiwavelength Astronomy Tsutomu T. TAKEUCHI. Suchetha](https://reader033.vdocuments.us/reader033/viewer/2022051807/6003f66893f6c563617421fe/html5/thumbnails/20.jpg)
The 2-D embedding recovered by Isomap.Isomap
Tenenbaum et al. (2011)
![Page 21: Machine Learning in the Era of Multiwavelength Astronomygopira.jp/sym2020/2-10-Takeuchi.pdf · Machine Learning in the Era of Multiwavelength Astronomy Tsutomu T. TAKEUCHI. Suchetha](https://reader033.vdocuments.us/reader033/viewer/2022051807/6003f66893f6c563617421fe/html5/thumbnails/21.jpg)
McInnes et al. (2018)
UMAP: Uniform Manifold Approximation and Projection forDimension Reduction
The algorithm is founded on three :
1. The data are uniformly distributed on a Riemannianmanifold.
2. The Riemannian metric is locally constant (or can beapproximated as such).
3. The manifold is locally connected.
From these assumptions it is possible to model the manifoldwith a fuzzy topological structure.
UMAP
![Page 22: Machine Learning in the Era of Multiwavelength Astronomygopira.jp/sym2020/2-10-Takeuchi.pdf · Machine Learning in the Era of Multiwavelength Astronomy Tsutomu T. TAKEUCHI. Suchetha](https://reader033.vdocuments.us/reader033/viewer/2022051807/6003f66893f6c563617421fe/html5/thumbnails/22.jpg)
Isomap preserves the density of the data point cloud (i.e.,distribution of galaxies) on the manifold (i.e. dense regionsremain dense).
The arrows show the continuous evolution of less-massive,actively star-forming galaxies to massive, quiescent galaxies.
Result: galaxy manifold from Isomap
![Page 23: Machine Learning in the Era of Multiwavelength Astronomygopira.jp/sym2020/2-10-Takeuchi.pdf · Machine Learning in the Era of Multiwavelength Astronomy Tsutomu T. TAKEUCHI. Suchetha](https://reader033.vdocuments.us/reader033/viewer/2022051807/6003f66893f6c563617421fe/html5/thumbnails/23.jpg)
Result: galaxy manifold from UMAP
Cooray et al. (2020)
![Page 24: Machine Learning in the Era of Multiwavelength Astronomygopira.jp/sym2020/2-10-Takeuchi.pdf · Machine Learning in the Era of Multiwavelength Astronomy Tsutomu T. TAKEUCHI. Suchetha](https://reader033.vdocuments.us/reader033/viewer/2022051807/6003f66893f6c563617421fe/html5/thumbnails/24.jpg)
Blue cloud: actively star formingRed sequence: quiescent Green valley: transient Cooray et al. (2020)
Result: galaxy manifold from UMAP
![Page 25: Machine Learning in the Era of Multiwavelength Astronomygopira.jp/sym2020/2-10-Takeuchi.pdf · Machine Learning in the Era of Multiwavelength Astronomy Tsutomu T. TAKEUCHI. Suchetha](https://reader033.vdocuments.us/reader033/viewer/2022051807/6003f66893f6c563617421fe/html5/thumbnails/25.jpg)
UMAP expands dense regions and contracts sparseregions, i.e., uniform representation of the observationalgalaxy parameters.
Result: galaxy manifold from UMAP
The arrows show the continuous evolution of less-massive,actively star-forming galaxies to massive, quiescent galaxies.
![Page 26: Machine Learning in the Era of Multiwavelength Astronomygopira.jp/sym2020/2-10-Takeuchi.pdf · Machine Learning in the Era of Multiwavelength Astronomy Tsutomu T. TAKEUCHI. Suchetha](https://reader033.vdocuments.us/reader033/viewer/2022051807/6003f66893f6c563617421fe/html5/thumbnails/26.jpg)
1. We demonstrated how machine learning methods canaid our understanding of galaxy evolution.
2. We found highly nonlinear continuous structure in themultidimensional luminosity space: the galaxy manifold.This is a revived, improved version of the conceptdiscussed in 80’s.
3. All the known empirical relations are projection of themanifold.
4. The galaxy manifold represents the evolutionarysequence of galaxies. Possibly we can parametrize thegalaxy evolution by a few parameters.
4. Conclusions
![Page 27: Machine Learning in the Era of Multiwavelength Astronomygopira.jp/sym2020/2-10-Takeuchi.pdf · Machine Learning in the Era of Multiwavelength Astronomy Tsutomu T. TAKEUCHI. Suchetha](https://reader033.vdocuments.us/reader033/viewer/2022051807/6003f66893f6c563617421fe/html5/thumbnails/27.jpg)
Take-home message
Forget about colors, and formulate your problem in luminosity(and other feature) space!
Machine learning can serve as a powerful and revolutionarymethod to develop fundamental part of astrophysics!
4. Conclusions