(Astro)physics-at-scale,-without-(Astro)physicists

Joshua'Bloom' UC'Berkeley,'Astronomy

@pro%sb POTUS-2014-2-NYC2-21-Sep-2014 Binary deeply interesting: Reveal the fundamental properties of stars

Harvard College Observatory c. 1890 50k-variables,-810-with-known-labels-(Dmeseries,-colors) (Semi)-Supervised-Machine:Learning-Approach-to-Classifica?on

Engineered-features!homogenize-data-→-p Describe-Dme2domain-characterisDcs-&-context-of-a-source

p-≈-100-features-computed-in-<-1-sec-per-82core-machine-(including-periodogram- analysis)

variability-metrics: e.g.-dominant-frequencies-in-Lomb2Scargle,-periodic-metrics: e.g.-Stetson-indices,-χ2/dof-(constant- phase-offsets-between-periods hypothesis)

shape-analysis context-metrics e.g.-distance-to-nearest-,-type-of- e.g.-skewness,-kurtosis,-Gaussianity nearest-galaxy,-locaDon-in-the-eclipDc-plane

Wózniak-et-al.-2004;-Protopapas+06,-Willemsen-&-Eyer-2007;-Debosscher-et-al.-2007;-Mahabal-et-al.-2008;-Sarro-et-al.-2009;-Blomme-et-al.- 2010;-Kim+11,-Richards+11 Machine!Learned!Classifica0onThe Astrophysical Journal Supplement Series,203:32(27pp),2012December Richards et al. True Class

a b1 b2 b3 b4 c d e f g h i j j1 l o p q r1 s1 s2 s3 t u v w x y

a. Mira 0.923 0.015 0.042 0.057 0.176

b1. Semireg PV 0.066 0.824 0.276 0.111 0.125 0.086 0.25 0.059 0.375

b2. SARG A 0.6 0.034 0.019 0.059

b3. SARG B 0.015 0.267 0.586 0.25 0.05 1

b4. LSP 0.011 0.074 0.067 0.069 0.87 0.2 0.171 0.059

c. RV Tauri 0.029 0.75 0.011 0.059 0.125 0.015

d. Classical Cepheid 0.955 0.538 0.25 0.25 0.1

e. Pop. II Cepheid 0.042 0.308 0.25

f. Multi. Mode Cepheid 0.011 0.077 0.5 0.012 0.043 0.1 0.059 1 g. RR Lyrae, FM 0.011 0.077 0.965 0.5 1 74- h. RR Lyrae, FO 0.012 0.897 0.036 0.015

i. RR Lyrae, DM 0.5 j. Delta Scuti 0.034 0.786 0.278 0.015 dimensional- j1. SX Phe l. Beta Cephei 0.107 0.722 0.015 feature-set-for- o. Pulsating Be 0.067 0.4 0.059

p. RSG 0.044 0.686 0.125 Predicted Class q. Chem. Peculiar 0.036 0.2 0.913 0.015 learning,-25- r1. RCB 0.706 s1. Class. T Tauri classes s2. Weak−line T Tauri 0.034 0.011 0.043 0.7 0.118 0.061

s3. RS CVn 0.05 0.059 0.037

t. Herbig AE/BE 0.2 0.375

u. S Doradus

v. Ellipsoidal

w. Beta Persei 0.353 0.889 0.182

x. Beta Lyrae 0.059 0.118 0.074 0.667 0.044

y. W Ursae Maj. 0.042 0.012 0.069 0.036 0.25 0.059 0.091 0.882

91 68 15 29 54 24 89 13 4 86 29 2 28 1 18 5 35 23 17 4 20 17 8 1 1 27 33 68 Figure 5. Cross-validated confusion matrix for all 810 ASAS training sources. Columns are normalized to sum to unity, with the total number of true objects of each Richards+12 class listed along the bottom axis. The overall correspondence rate for these sources is 80.25%, with at least 70% correspondence for half of the classes. Classes with low correspondence are those with fewer than 10 training sources or classes which are easily confused. Red giant classes tend to be confused with other red giant classes and eclipsing classes with other eclipsing classes. There is substantial power in the top-right quadrant, where rotational and eruptive classes are misclassified as red giants; these errors are likely due to small training set size for those classes and difficulty to classify those non-periodic sources. (A color version of this figure is available in the online journal.)

with any monotonically increasing function (which is typically probabilities, pi1, pi2,...,piC ,areproperprobabilitiesinthat restricted to a set of non-parametric isotonic functions, such as they are each{ between 0 and 1 and} sum to unity for each object. step-wise constants). A drawback to both of these methods is The optimal value of r is found by minimizing the Brier score ! ! ! that they assume a two-class problem; a straightforward way (Brier 1950)betweenthecalibrated(cross-validated)andtrue around this is to treat the multi-class problem as C one-versus- probabilities.14 We find that using a fixed value for r is too all classification problems, where C is the number of classes. restrictive and, for objects with small maximal RF probability, However, we find that Platt Scaling is too restrictive of a trans- it enforces too wide of a margin between the first- and second- formation to reasonably calibrate our data and determine that largest probabilities. Instead, we implement a procedure similar we do not have enough training data in each class to use Isotonic to that of Bostrom (2008)andparameterizer with a sigmoid Regression with any degree of confidence. function based on the classifier margin, ∆i pi,max pi,2nd, for Ultimately, we find that a calibration method similar to the each source, = − one introduced by Bostrom (2008)isthemosteffectiveforour 1 1 data. This method uses the probability transformation r(∆i) , (5) = 1+eA∆i +B − 1+eB

pij + r(1 pij )ifpij max pi1,pi2,...,piC where the second term ensures that there is zero calibration per- pij − = { } formed at ∆i 0. This parameterization allows the amount of = pij (1 r)otherwise, = " − (4) 14 N C 2 The Brier score is defined as B(p) 1/N (I(yi j) pij ) , = i 1 j 1 = − where! pi1,pi2,...,piC is the vector of class probabilities where N is the total number of objects, C is the number= = of classes, and { } # # for object i and r [0, 1] is a scalar. Note that the adjusted I(yi j) is 1 if and only if the true class of the source i is j. ∈ = ! ! 11 Machine8learned'Variable''catalog: hDp://bigmacc.info Doing-Science-with-Probabilis?c-Catalogs

Demographics-(with-ligle-followup): ------trading-high-purity-at-the-cost-of-lower-efficiency !!!!!!!e.g.,!using!RRL!to!find!new!Galac5c!structure

Novelty-Discovery-(with/lots/of/followup): ------trading-high-efficiency-for-lower-purity !!!!!!!e.g.,!discovering!new!instances!of!rare!classes!! –13–

DRAFT April 20, 2012

Discovery of Bright Galactic and DY Persei Variables: Rare Gems Mined from ASAS

1, 1,2 1 1 1 A. A. Miller ⇤, J. W. Richards , J. S. Bloom , S. B. Cenko , J. M. Silverman , D. L. Starr1, and–13– K. G. Stassun3,4

ABSTRACT

We present the results of a machine-learning (ML) based search for new R Coronae Borealis (RCB) stars and DY Persei-like stars (DYPers) in the Galaxy using cataloged light curves obtained by the All-Sky Automated Survey (ASAS). RCB stars—a rare class of hydrogen-deficient carbon-rich supergiants—are of great interest owing to the insights they can provide on the late stages of stellar evolution. DYPers are possibly the low-temperature, low-luminosity analogs to the RCB phenomenon, though additional examples are needed to fully estab- lish this connection. While RCB stars and DYPers are traditionally identified by epochs of extreme dimming that occur without regularity, the ML search framework more fully captures the richness and diversity of their photometric behavior. We demonstrate that our ML method recovers ASAS candidates that would have been missed by traditional search methods employing hard cuts on amplitude and periodicity. Our search yields 13 candidates that we consider likely RCB stars/DYPers: new and archival spectroscopic observations confirm that four of these candidates are RCB stars and four are DYPers. Our discovery of four new DYPers increases the number of known Galactic DYPers from two to six; noteworthy is that one of the new DYPers has a measured parallax and is m 7 mag, making it the brightest known DYPer to date. Future observations ⇡ of these new DYPers should prove instrumental in establishing the RCB con- arXiv:1204.4181v1 [astro-ph.SR] 18 Apr 2012 nection. We consider these results, derived from a machine-learned probabilistic

Fig. 2.— ASAS V -band1Department light of Astronomy, curves University of of California,newly Berkeley, discovery CA 94720-3411, RCBUSA stars and DYPers. Note the di↵17-known-GalacDc-RCB/DY-Perering magnitude ranges2Statistics Department, shown University for of each California, Berkeley,light CA, curve. 94720-7450, USA Spectroscopic observations confirm 3Department of Physics and Astronomy, Vanderbilt University, Nashville, TN 37235, USA the top four candidates4Department to be of Physics, RCB Fisk University, stars, 1000 while 17th Ave. N., the Nashville, bottom TN 37208, USA four are DYPers. *E-mail: [email protected]

Fig. 2.— ASAS V -band light curves of newly discovery RCB stars and DYPers. Note the di↵ering magnitude ranges shown for each light curve. Spectroscopic observations confirm the top four candidates to be RCB stars, while the bottom four are DYPers. Turning-Imagers-into-Spectrographs Time/variability/+/colors/→/fundamental/stellar/parameters

Data:-5000-variables-in-SDSS-Stripe-82-with-spectra ---~80-dimensional-regression-with-Random-Forest Miller,/JSB+14 Extragalac?c-Transient-Universe:-Explosive-Systems

-22

-20 Pair Production Supernovae z=0.45

-18 Type Ia H

M -16 Type IIp

-14 IMBH + WD Collision 200Mpc NS + RSG Collision -12 -log(brightness) NS + NS Mergers -10 50 100 150 200 Days Since Explosion E. Ramirez-Ruiz (UCSC)

Towards-a-Fully-Automated-Stack-for-Time:Domain-Transients

current papers state)of)the)art } inference stack followup classificaDon discovery finding reducDon observing scheduling automated-(e.g.,-iPTF)- strategy not-(yet)-automated

1106.5491 Op?mize-scien?fic-return-in-a-resource:aware-context,-without-people

Palomar-Transient-Factory- Zwicky-Transient-Factory- Large-Synop?c-Survey-Telescope- (PTF) (ZTF) (LSST) 200922016 201722020 202022030

Image-data-rate 1/GB/90s 3/GB/45/s 6/GB/5/s

Transient- ✕ 4 ✕ 5 ✕ 6 Alerts-per-night 4 10 3 10 2 10 “cheap”-discovery

“expensive”-followup

Hubble-Space-Telescope-(HST) James-Webb-Space-Telescope-(JWST) Thirty2Meter-Telescope-(TMT) Smith+11 Smith+11 Machine Learning the Supernova Zoo Joseph W. Richards* and Joshua S. Bloom Department of Astronomy, University of California Berkeley *[email protected]

Abstract Predicting Supernova Zoo Scores with Real-Time ML Finding Supernovae

In the SN Zoo, every candidate SN is vetted by Our goal is to find new supernovae in imaging data. Identifying supernovae in terabytes of imaging Supernova several users. The candidate is assigned, by each Asteroid We find that using the RF score to select SNe produces 1.000 data is a non-trivial task. Recently, the Galaxy user, a score in {-1,1,3} with 3 signifying a likely SN. Junk/Unconfirmed samples of higher efficiency and purity than using the Zoo Supernovae project [1] was unveiled as a We train a Random Forest classifier on real-time SN Zoo score. 0.500 features to automatically predict which objects citizen-science approach to supernova (SN) Classification of known SNe identification. An invaluable by-product of will attain a median Zoo score of 3. these Zoo efforts is an immense amount of Figure 1. ROC curve for RF prediction of which 0.5 Random Forest sources will have median Zoo score of 3. At a Supernova Zoo data from which to train automated methods for 0.100 10% (5%) false positive rate, the classifier only supernova discovery. We use the first year of 0.4 misses 13% (19%) of all candidates whose true 0.050 data from Galaxy Zoo Supernovae to train median Zoo score is 3. The red envelope is machine learning (ML) algorithms to predict 90% confidence band. 0.3 the Zoo consensus on each object. Using 0.5 Classification of Candidates image processing and contextual features, we 0.2 Machine2learned,-immediate- Random Forest Score Random Forest w/ Median Zoo Score = 3 0.010 find that, in real time, we can predict which

0.4 classificaDon-out2performs 0.005 objects will be highly rated by the Zoo with 0.1 Missed Detection Rate (MDR) 90% efficiency and 87% purity. Furthermore, crowdsourced-SN-discovery we discover that, in terms of discovering SNe, 0.3 0.0 19% of Zoo 3's our ML method attains a 14.5% missed 0.0 0.1 0.2 0.3 0.4 0.5 missed at 5% FPR 0.001

0.2 ● False Positive Rate (FPR) detection rate at 10% false positive rate, −1 0 1 2 3 13% of Zoo 3's missed at 10% FPR compared to a much higher missed detection ● Mean Zoo Score Figure 4. By lowering the SN selection threshold, we Missed Detection Rate (MDR) rate of 34.2% attained by using the Zoo score. 0.1 trade lower MDR for higher FPR. For a sample of 345 Figure 2. Cross-validated RF score (PrRF(Zoo = 3)) versus the mean known PTF SNe, employing the RF score to select Zoo score for all 25,913 candidates. Objects with larger Zoo score objects is uniformly better, in terms of MDR and FPR, typically have high RF score. Most of the confirmed supernovae (red 0.0 than using the SN Zoo score. At 10% FPR, the RF Data and Methods triangles) and asteroids (cyan stars) have RF prob. > 0.1. There are a 0.0 0.1 0.2 0.3 0.4 0.5 criterion (threshold = 0.035) attains a 14.5% MDR number of non-SNe with large mean Zoo Score and low RF score. False Positive Rate (FPR) compared to 34.2% MDR for SN Zoo (threshold = 0.4). Data: We use imaging data taken by the Palomar Transient Richards+13- Factory (PTF) on 25913 supernova candidates. Of the Candidates with Low Zoo Score & High RF Score candidates, 345 are spectroscopically confirmed Conclusions supernovae. Each object was vetted by the SN Zoo. Figure 3a. PTF10gop, Objects with lowest possible median Zoo score (-1) The Galaxy Zoo Supernovae project is a powerful tool for a type II SN with good and high Random Forest score (sorted by RF score): SN discovery. We show that data collected from the Zoo Features: image subtraction but can be employed to effectively train machine learning Each candidate is represented by a set of features that can LBL ID PTF name Type Mean Zoo score RF score anomalously low SN algorithms to automatically detect supernovae in be computed at the time of detection. 217987666 10pme nova? 0.238 0.604 Zoo score. Image processing features - include magnitude (of 122041291 cosmic ray 0.619 0.564 massive data streams. In future work, we will use ML to 201381225 asteroid 0.166 0.518 improve the Zoo w/ better candidate selection and scoring. reference and subtraction image), the fit of a 2-D Gaussian 326421262 cosmic ray 0.166 0.518 Figure 3b. PTF10vty, a to the subtraction image, and real-bogus score [2]. 321279994 10vty type Ia SN 0.714 0.509 type Ia SN. Source Context features - include SDSS and USNO info, redshift, 156393144 asteriod 0.166 0.502 subtracts well despite 283389856 asteroid? -0.090 0.500 and proximity to nearest galaxy. We use 62 features for diffraction spike in References each source. See [2] for more details. 154615556 10hvz type II SN 0.523 0.488 121171477 asteroid 0.523 0.484 reference image. [1] Smith, et al., (2010) Galaxy Zoo Supernovae MNRAS 412, 1309 162859292 asteroid 0.142 0.472 [2] Bloom, et al. (2011) Automating Discovery and Classification of Transients and Classifier: 248639054 asteroid -0.400 0.466 Figure 3c. PTF10pme, Variable Stars in the Synoptic Survey Era, arXiv:1106.5491, submitted to PASP We use a Random Forest (RF) [3,4] classifier to learn the 260838581 asteroid 0.333 0.462 [3] Brieman (2001) Random Forests, Machine Learning 45, 5 a real source (nova?) w/ mapping from the features to the median SN Zoo score. RF 138363313 10gop type Ia SN 0.090 0.440 [4] Richards, et al., (2011) On Machine-learned Classification of Variable Stars builds a set of de-correlated decision trees on bootstrapped 119552814 asteroid 0.090 0.414 low Zoo score, perhaps with Sparse and Noisy Time-series Data, ApJ 733, 10 194434389 unknown -0.600 0.406 due to crowded field and samples from a training set & averages the trees to yield a We acknowledge the generous support of a Cyber-Enabled Discovery and robust, low-bias, low-variance classification model. Table 1: Median Zoo = -1 objects with highest RF Scores residuals of nearby stars Innovation (CDI) grant (No. 0941742) from the NSF. This work was performed in the CDI-sponsored Center for Time Domain Informatics (http://cftd.info).

1 PTF11kly((SN(2011fe)

©Peter&Nugent

Supernova-Discovery-in-the-Pinwheel-Galaxy ////11/hr/aNer/explosion nearest-SN-Ia-in->3-decades Discovered'by'our'machine8learning'framework in-PTF:->10,000-events-in->-0.2-PB-of-imaging-→-50+-journal-arDcles –13–

Also,/Li,/JSB/et/al./Nature,/2011 ////////Nugent...JSB,/Nature,/2011

The Astrophysical Journal Letters,744:L17(5pp),2012January10 doi:10.1088/2041-8205/744/2/L17 C 2012. The American Astronomical Society. All rights reserved. Printed in the U.S.A. Fig. 3.—⃝ Constraints on mass, effective temperature, radius and average density of the primary star of SN 2011fe. The shaded red region is excluded from non-detection of an optical quiescent counterpartA COMPACT DEGENERATE in Hubble PRIMARY-STAR Space Telescope PROGENITOR (HST) OF SN imaging. 2011fe The shaded green Joshua S. Bloom1,2,DanielKasen1,3,4,KenJ.Shen1,3,12,PeterE.Nugent1,2,NathanielR.Butler5, region is excludedMelissa L. Graham from6, considerations7,D.AndrewHowell6,7 of,UlrichKolb the non-detection8,StefanHolmes8,CaroleA.Haswell of a shock breakout8, at early Vadim Burwitz9,JuanRodriguez10,andMarkSullivan11 times, taking the1 Department least of Astronomy, constraining University of California,Rp Berkeley,of the Berkeley three CA, 94720-3411, model USA; [email protected] Table 1. The blue region is 2 Physics Division, Lawrence Berkeley National Laboratory, Berkeley, CA 94720, USA excludes by the non detection3 Nuclear Science Division, of a Lawrence quiescent Berkeley National counterpart Laboratory, Berkeley, in CA 94720, Chandra USA X-ray imaging. The 4 Department of Physics, University of California, Berkeley, Berkeley CA, 94720, USA 5 School of Earth and Space Exploration, Arizona State University, P.O. Box 871404, Tempe, AZ 85287-1404, USA location of the H, He,6 Las Cumbres and Observatory C main Global sequence Telescope Network, is 6740 shown, Cortona Dr., with Suite 102, the Goleta, symbol CA 93117, USA size scaled for different 7 Department of Physics, University of California Santa Barbara, Santa Barbara, CA 93106, USA primary masses. Several8 Department observed of Physical Sciences, WD The Open and University, NSs Walton are Hall, shown. Milton Keynes MK7 The 6AA, primary UK radius in units of 9 Max-Planck-Institut fur¨ extraterrestrische Physik, Giessenbachstrasse, 85748 Garching, Germany 10 Observatori Astronomic` de Mallorca, Cam´ı de l’Observatori, 07144 Costitx, Mallorca, Spain R is shown for Mp =111 .4M . ⊙ Department of⊙ Physics (Astrophysics), University of Oxford, Keble Road, Oxford OX1 3RH, UK Received 2011 October 17; accepted 2011 November 28; published 2011 December 15

ABSTRACT While a white dwarf (WD) is, from a theoretical perspective, the most plausible primary star of a Type Ia supernova (SN Ia), many other candidates have not been formally ruled out. Shock energy deposited in the envelope of any exploding primary contributes to the early SN brightness and, since this radiation energy is degraded by expansion after the explosion, the diffusive luminosity depends on the initial primary radius. We present a new non-detection limit of the nearby SN Ia 2011fe, obtained at a time that appears to be just 4 hr after explosion, allowing us to directly constrain the initial primary radius (Rp). Coupled with the non-detection of a quiescent X-ray counterpart and the 56 inferred synthesized Ni mass, we show that Rp ! 0.02 R (a factor of five smaller than previously inferred), that 4 ⊙ 3 the average density of the primary must be ρp > 10 g cm− ,andthattheeffectivetemperaturemustbelessthan afew 105 K. This rules out hydrogen-burning main-sequence stars and giants. Constructing the helium-burning and carbon-burning× main sequences, we find that such objects are also excluded. By process of elimination, we find that only degeneracy-supported compact objects—WDs and neutron stars—are viable as the primary star of SN 2011fe. With few caveats, we also restrict the companion (secondary) star radius to Rc ! 0.1 R ,excluding Roche-lobe overflowing red giant and main-sequence companions to high significance. ⊙ Key words: supernovae: individual (2011fe, white dwarfs) – supernovae: general

1. INTRODUCTION (3) degenerate objects can result in runaway, explosive burning, (4) the energy gained from burning a WD matches that seen in While the nature of the explosion that leads to a Type Ia an SN Ia, and (5) simulations of the explosions of C+O WD supernova (SN Ia)—detonation (Woosley et al. 1986;Livne stars are successful at reproducing SN Ia spectra (e.g., Nomoto &Arnett1995), deflagration (Nomoto et al. 1976), or both 1982). However, the discovery of a class of SNe Ia that require (Khokhlov 1991)—and the process that leads to the explosion a WD mass above the Chandrasekhar limit (Howell et al. 2006) trigger are not well known, it is commonly assumed that an SN Ia has caused some to question whether a WD is involved in the is powered by the explosion of a white dwarf (WD) at a pressure explosion after all (Taubenberger et al. 2011). and temperature sufficient to ignite carbon (Nomoto 1982;Iben The SN Ia SN 2011fe was discovered in the Pinwheel &Tutukov1984). All viable SN Ia models of the progenitor Galaxy (M101) more than two weeks before it hit maximum system include a companion (secondary) star that transfers mass brightness on 2011 September 12 UT (Nugent et al. 2011). To (either steadily or violently) to the WD (Nomoto et al. 1997; date, SN 2011fe provides some of the best constraints on the Podsiadlowski et al. 2008). Single-degenerate channels involve progenitor system of an SN Ia: coupled with a well-measured the transfer of mass from a giant, , or helium star distance modulus to M101 (DM 29.05 0.23 mag; Shappee (Whelan & Iben 1973;Nomoto1982;Munari&Renzini1992; &Stanek2011), the non-detection= of a quiescent± counterpart at Han & Podsiadlowski 2004). Double-generate channels involve optical, infrared, mid-infrared, and X-ray wavebands were used the merger of two WDs (Webbink 1984;Iben&Tutukov1984). to placed strict limits on the nature of the progenitor system. There is considerable circumstantial evidence that the pri- With optical imaging reaching 100 times fainter than previous mary is a C+O WD (Hoyle & Fowler 1960), but until recently limits, Li (2011), Nugent et al.∼ (2011), and Horesh et al. (2011) (Nugent et al. 2011; see also Brown et al. 2011 for a similar showed that Roche-lobe overflow red giants as the secondary analysis) there have been very few direct constraints. The ev- star were excluded; He-star + WD progenitor systems were idence is (1) neither hydrogen nor helium are seen in SNe Ia also largely excluded. Nugent et al. (2011)placedconstraints (Leonard 2007), and few astrophysical objects lack these ele- on double-degenerate models based on the non-detection of ments, (2) the elements synthesized in SN Ia are consistent with early-time emission from shock interaction with the disrupted the fusion chain leading from carbon going up to the iron peak, secondary WD material (Shen et al. 2011). Rather than focus on the progenitor system as a whole, in this 12 Einstein Fellow. Letter we investigate what can be gleaned about the primary

1 Real2Dme-ClassificaDons... Real2Dme-ClassificaDons... 12 Singer et al.

Figure 10. Localization of a typical circa 2016 GW detection in the HLV network configuration. This is a Mollweide projection in geographic coordinates. This 2 event consists of triggers in H and L and has a SNRnet = 13.4. The rapid sky localization gives a 90% confidence region with an area of 1170 deg and the full stochastic sampler gives 515 deg2. This is event ID 821759 from Tables 4 and 5 and the online supplement (see Appendix for more details). (A color version of this figure is available in the online journal.) Applica'on:+Finding+the+first+Electromagne'c+Counterpart+of+a+Gravity+Wave+Event

Double/Neutron/star/Merger Singer+2014

in-50-deg2-(≈250-full-moons): •-107--(103-nearby) (a) BAYESTAR •--~dozens-of-ongoing-supernovae •--hundreds-of-flaring-stars-

Early demonstration: Discovery & redshift of a gamma-ray burst optical afterglow in 71 deg2 Singer,-JSB-+2014-arXiv:1307.5851

(b) LALINFERENCE_MCMC Can-we-infer-a-physical-taxonomy-&-find-new-classes?

?

Survey Data Taxonomy The Astrophysical Journal Supplement Series,203:32(27pp),2012December Richards et al.

True Class large a b1 b2 b3 b4 c d e f g h i j j1 l o p q r1 s1 s2 s3 t u v w x y a. Mira 0.923 0.015 0.042 0.057 0.176 pulsational b1. Semireg PV 0.066 0.824 0.276 0.111 0.125 0.086 0.25 0.059 0.375 b2. SARG A 0.6 0.034 0.019 0.059

b3. SARG B 0.015 0.267 0.586 0.25 0.05 1

b4. LSP 0.011 0.074 0.067 0.069 0.87 0.2 0.171 0.059

c. RV Tauri 0.029 0.75 0.011 0.059 0.125 0.015

d. Classical Cepheid 0.955 0.538 0.25 0.25 0.1

e. Pop. II Cepheid 0.042 0.308 0.25

f. Multi. Mode Cepheid 0.011 0.077 0.5 0.012 0.043 0.1 0.059 1

g. RR Lyrae, FM 0.011 0.077 0.965 0.5 1

h. RR Lyrae, FO 0.012 0.897 0.036 0.015

i. RR Lyrae, DM 0.5

j. Delta Scuti 0.034 0.786 0.278 0.015

j1. SX Phe

l. Beta Cephei 0.107 0.722 0.015

o. Pulsating Be 0.067 0.4 0.059

p. RSG 0.044 0.686 0.125

Predicted Class q. Chem. Peculiar 0.036 0.2 0.913 0.015

r1. RCB 0.706

s1. Class. T Tauri

s2. Weak−line T Tauri 0.034 0.011 0.043 0.7 0.118 0.061

s3. RS CVn 0.05 0.059 0.037

t. Herbig AE/BE 0.2 0.375

u. S Doradus v. Ellipsoidal multistar w. Beta Persei 0.353 0.889 0.182

x. Beta Lyrae 0.059 0.118 0.074 0.667 0.044

y. W Ursae Maj. 0.042 0.012 0.069 0.036 0.25 0.059 0.091 0.882

91 68 15 29 54 24 89 13 4 86 29 2 28 1 18 5 35 23 17 4 20 17 8 1 1 27 33 68 Figure 5. Cross-validated confusion matrix for all 810 ASAS training sources. Columns are normalized to sum to unity, with the total number of true objects of each class listed along the bottom axis. The overall correspondence rate for these sources is 80.25%, with at least 70% correspondence for half of the classes. Classes with low correspondence are those with fewer than 10 training sources or classes which are easily confused. Red giant classes tend to be confused with other red giant classes and eclipsing classes with other eclipsing classes. There is substantial power in the top-right quadrant, where rotational and eruptive classes are misclassified as red giants; these errors are likely due to small training set size for those classes and difficulty to classify those non-periodic sources. (A color version of this figure is available in the online journal.)

with any monotonically increasing function (which is typically probabilities, pi1, pi2,...,piC ,areproperprobabilitiesinthat restricted to a set of non-parametric isotonic functions, such as they are each{ between 0 and 1 and} sum to unity for each object. step-wise constants). A drawback to both of these methods is The optimal value of r is found by minimizing the Brier score ! ! ! that they assume a two-class problem; a straightforward way (Brier 1950)betweenthecalibrated(cross-validated)andtrue around this is to treat the multi-class problem as C one-versus- probabilities.14 We find that using a fixed value for r is too all classification problems, where C is the number of classes. restrictive and, for objects with small maximal RF probability, However, we find that Platt Scaling is too restrictive of a trans- it enforces too wide of a margin between the first- and second- formation to reasonably calibrate our data and determine that largest probabilities. Instead, we implement a procedure similar we do not have enough training data in each class to use Isotonic to that of Bostrom (2008)andparameterizer with a sigmoid Regression with any degree of confidence. function based on the classifier margin, ∆i pi,max pi,2nd, for Ultimately, we find that a calibration method similar to the each source, = − one introduced by Bostrom (2008)isthemosteffectiveforour 1 1 data. This method uses the probability transformation r(∆i) , (5) = 1+eA∆i +B − 1+eB

pij + r(1 pij )ifpij max pi1,pi2,...,piC where the second term ensures that there is zero calibration per- pij − = { } formed at ∆i 0. This parameterization allows the amount of = pij (1 r)otherwise, = " − (4) 14 N C 2 The Brier score is defined as B(p) 1/N (I(yi j) pij ) , = i 1 j 1 = − where! pi1,pi2,...,piC is the vector of class probabilities where N is the total number of objects, C is the number= = of classes, and { } # # for object i and r [0, 1] is a scalar. Note that the adjusted I(yi j) is 1 if and only if the true class of the source i is j. ∈ = ! ! 11 Unsupervised-Learning:-Finding-Anomalous-Classes...- HenrieDa-Swan-LeaviD

Discoverer-of-the-period2 luminosity-relaDon-in- Cepheid-stars ‣-Basis-of-Hubble’s- expansion-of-the-universe ‣-Genesis-of-modern- cosmology HenrieDa-Swan-LeaviD

Discoverer-of-the-period2 Can!machines!be!taught!to!ask!ques0ons!we!haven’t!luminosity-relaDon-in- or!cannot?!!Cepheid-starsWill!machine!intelligence!ever!replace!the! Eureka!!moments!by!people?‣-Basis-of-Hubble’s- !Will!astrophysics!ever!be! expansion-of-the-universetruly!automated? ‣-Genesis-of-modern- cosmology “The End of Theory: The Data Deluge Makes the Scientific Method Obsolete” by Chris Anderson

http://archive.wired.com/science/discoveries/magazine/16-07/pb_theory# Some may become more Extremely significant Censored with time... And many of these are correct, or specious methods, or fraud.

Some less...

Kühberger A, Fritz A, Scherndl T (2014) Publication Bias in Psychology: A Diagnosis Based on the Correlation between Effect Size and Sample Size. PLoS ONE 9(9): e105825. doi:10.1371/journal.pone.0105825 http://www.plosone.org/article/info:doi/10.1371/journal.pone.0105825 SCIENCE VOL 324 3 APRIL 2009 MachineCassisted! “discovery”!of the!fundamental! plane!of!galaxies

Predicts −5/4 1/4 -0.07 Original (87) log(re) = log σ + 0.27<µ> − 4.1 <µ> ∼ L ; Dn ∝ σ0 <µ>

−3 0.16 Assisted (11) log(re) = 1.4 log σ + 0.36<µ> − 6.7 <µ> ∼ L ; Dn ∝ σ0 <µ> most&parsimonious&/&same&RMS&as&original Graham+11 Par'ng+Thoughts • Astronomy’s-data-deluge-demands-an-abstracDon-of-the- tradiDonal-roles-in-the-scienDfic-process.-

• Machine-Learning-(&-Parallel,-Distributed-CompuDng)-and-at- the-heart-of-what-we-do

• But-Machine-Learning-is-only-truly-useful-in-service,-enabling- novel-science

• With-unsupervised-learning,-we’re-taking-baby-steps-towards- inference Thanks!

@pro%sb –12–

Fig. 2.— Absolute g-band magnitude versus time since explosion in three theoretical models for the early time evolution of Type Ia SNe. Shown is 4 hr, 5 σ non-detection discussed in 2.2 and the first two detections from PTF (Nugent et al. 2011). The black line shows § the L t2 behavior of the early-time light curve detected in PTF, consistent with the non- ∝ detection. For the Kasen (2010) companion interaction model, R denotes the separation distance between the two stars, and the light curve is shown for an observer aligned with the collision axis, which produces the brightest observed luminosity. Constraint on Secondary

assuming Roche Lobe overflow, Rs ≲ a/2

for unfavorable viewing angles, observed flux from ejecta interaction may be ~10 smaller than optimum:

Rs ≲ 0.1 R⦿

apparently excludes MS & RGs Constraining the Radius

2 1. Temp-Radius Lν∝ BB(T0) × Rp X-ray/optical preimaging see also, Liu, di Stefano+12 2. Shock Breakout detonation➝shock Piro, Chang, Weinberg 10 ➝ early-time emission Rabinak & Waxman 11 40 -1/3 Lbol ~ 10 erg/s × Rp tday 1/4 -1/2 T(t)~ 4000 K × Rp tday Constraining the Radius

2 1. Temp-Radius Lν∝ BB(T0) × Rp X-ray/optical preimaging see also, Liu, di Stefano+12 2. Shock Breakout detonation➝shock Piro, Chang, Weinberg 10 ➝ early-time emission Rabinak & Waxman 11 40 -1/3 Lbol ~ 10 erg/s × Rp tday 1/4 -1/2 T(t)~ 4000 K × Rp tday

3. Ejecta heating close-in secondary gets heated secondary by SN ejecta Kasen 11 42 -1/2 Lbol ~ 7×10 erg/s × aAU tday 1/4 -37/72 T(t)~ 4125 K × aAU tday