The changing landscape of astroinformatics

Transcription

1 The changing landscape of astroinformatics astrostatistics and Eric Feigelson Center for Astrosta2s2cs Penn State University IAU Symposium #325 Astroinforma6cs Sorrento IT

2 An aphorism The scien)st collects data in order to understand natural phenomena The sta)s)cian helps the scien)st acquire understanding from the data The computer scien)st helps the scien)st perform the needed calcula)ons (*) The individual proficient at all these tasks is a data scien)st (*) needed only for Big Data or Big Theory

3 Outline 1 Astronomers & stat/comp methodology: o Role of statistics in astronomy o Education gap o Resurgence of astrostatistics 2 The R statistical software environment

4 Statistical questions for a scientist How reliable are my results? (P=0.05 vs. P=0.003) Have I used the most effective procedures for achieving my results? (KS vs. AD test) Do my results depend on uncertain models? (power law?) Do my results depend on my chosen analysis path? A suggested path: Ø Nonparametric data analysis (including hypothesis tests, local regression, Fourier or wavelet transforms, etc) Ø Clustering & classification (if dataset is inhomogeneous) Ø Parametric analysis: maximum likelihood with model selection Ø Bayesian analysis: weight likelihood with prior knowledge Ø Discuss: What have I learned from each step along the path?

5 Recommended steps in the statistical analysis of scientific data The application of statistics can reliably quantify information embedded in scientific data and help adjudicate the relevance of theoretical models. But this is not a straightforward, mechanical enterprise. It requires: Ø exploration of the data Ø careful statement of the scientific problem Ø model formulation in mathematical form Ø choice of statistical method(s) Ø calculation of statistical quantities Ø judicious scientific evaluation of the results Astronomers often do not adequately pursue each step

6 Astronomy involves many branches of modern statistics Cosmology Statistics Galaxy clustering Spatial point processes, clustering Galaxy morphology Galaxy luminosity fn Power law relationships Regression, mixture models Gamma distribution Pareto distribution Weak lensing morphology Geostatistics, density estimation Strong lensing morphology Shape statistics Strong lensing timing Time series with lag Faint source detection False Discovery Rate SN Ia & QSO lightcurves Nonstationary time series CMB spatial analysis Markov fields, ICA, etc ΛCDM parameters Bayesian inference & model selection Comparing data & simulation Uncertainty quantification

7 The education gap Statistics Astronomers usually take 0-1 courses in statistical methods during their university/graduate education, in contrast to extensive training in physics and associated mathematics. Students are thirsty for statistics training. Computation Survey of ~1100 astronomers worldwide (Momcheva & Tollerud 2015) shows that 90% write software but only 8% receive substantial training in software development and 43% receive no training. Preferred languages are Python (67%), IDL (44%), C/C++ (37%), Fortran (28%),, Matlab (8%), R (6%, 3% in USA), julia (1%). No correlation with age. Considering the wide-spread use of R in other scientific fields, its popularity among astronomers is strikingly low. The result of weak training is a widespread ignorance of the myriad advances in modern statistical & computational methodology.

8 Consequently The state of astrostatistics today is often poor! Many astronomical studies are confined to a narrow suite of familiar statistical methods: Fourier transform for temporal analysis (Fourier 1807) Least squares regression (Legendre 1805, Pearson 1901) Kolmogorov-Smirnov goodness-of-fit test (Kolmogorov, 1933) Principal components analysis for tables (Hotelling 1936) Even traditional methods are sometimes misused see Beware the Kolmogorov-Smirnov test! page on ASAIP

9 Under-utilized methodology: modeling (MLE, EM Algorithm, BIC, bootstrap) multivariate classification (LDA, SVM, CART, RFs) time series (autoregressive models, state space models) spatial point processes (Ripley s K, kriging) nondetections (survival analysis) image analysis (computer vision methods, False Detection Rate) statistical computing (R) Advertisement Modern Statistical Methods for Astronomy with R Applications E. D. Feigelson & G. J. Babu, Cambridge Univ Press, 2012! "#$$%&!'()'!*+,-.!/01&2!34&!! 5%67!/67&4$489!:!;4684<4=9!544>!!

10 Recent resurgence in astrostatistics Papers in astronomical literature doubled to ~500/yr in past decade ( Methods: statistical papers in NASA-Smithsonian Astrophysics Data System) Short training courses (Penn State, India, Brazil, Greece, China, Italy, France, Germany, Spain, CfA, Caltech, Sweden, IAU/AAS/CASCA/ meetings) Cross-disciplinary research collaborations (Harvard/ICHASC, Carnegie-Mellon, Penn State, NASA-Ames/Stanford, CEA-Saclay/Stanford, Cornell, UC-Berkeley, Michigan, Imperial College London, Swinburne, ) Cross-disciplinary conferences (Statistical Challenges in Modern Astronomy, Astronomical Data Analysis , PhysStat, SAMSI 2006/2012, Astroinformatics ) Scholarly society working groups: Int l Stat Institute s Int l Astrostatistical Assn, Int l Astro Union Commission, Amer Astro Soc Working Group, Amer Stat Assn Interest Group, LSST Science Collaboration, IEEE Astro Data Miner Task Force

11 Textbooks New resources in astrostatistics Bayesian Logical Data Analysis for the Physical Sciences: A Compara:ve Approach with Mathema:ca Support Gregory, 2005 Prac:cal Sta:s:cs for Astronomers, Wall & Jenkins, 2 nd ed, 2012 Modern Sta:s:cal Methods for Astronomy with R Applica:on, Feigelson & Babu, 2012 Sta:s:cs, Data Mining, and Machine Learning in Astronomy: A Prac:cal Python Guide for the Analysis of Survey Data,, Ivecic, Connolly, VanderPlas & Gray, 2014 hop://asaip.psu.edu Recent papers, meetings, jobs, blogs, courses, forums,

12 Statistical analysis needs software A brief history of statistical computing 1960s c2000: Sta6s6cal analysis developed by academic sta6s6cians, but implemented by private companies (SAS, SPSS, etc) 1980s: John Chambers (ATT, USA)) develops S sta6s6cal package in the new C language 1990s: Ross Ihaka & Robert Gentleman (Univ Auckland NZ) mimic S in an open source system, R. R Core Development Team expands, GNU GPL release. 2000s: Comprehensive R Analysis Network (CRAN) for user- provided specialized packages grows exponen6ally. Important packages incorporated into base- R.

13 Growth of CRAN contributed packages Oct : 9373 packages (~5/day) ~200,000 functions See The Popularity of Data Analysis Software, R. A. Muenchen,

14 The R statistical computing environment R integrates data manipula6on, graphics and extensive sta6s6cal analysis. Uniform documenta6on and coding standards. Fully programmable C- like language, similar to IDL. Specializes in vector/matrix inputs. Same compiler speed as IDL, Python, Matlab. Easy binary download for Windows, Mac or linux. On- the- fly installa6on of CRAN packages. Quick 2- way communica6on with C, Fortran, Python. Emulator of Matlab user- provided add- on CRAN packages including ~20 for astronomy (FITSio, astrolibr, )

15 R s growing importance in data science Kdnuggests 2014 poll data analytics/mining SPSS R Posts on software forums 2013 Job trends from Indeed.com

16 R is now the world s premier statistical computing package developed by statisticians with community packages from all of the sciences extensively used in genomics, biostatistics, econometrics, ecology, social sci, etc Data scientists recommend both Python and R Usage of both is growing rapidly (

17 Some functionalities of base R arithme6c & linear algebra bootstrap resampling empirical distribu6on tests exploratory data analysis generalized linear modeling graphics robust sta6s6cs linear programming local and ridge regression max likelihood es6ma6on multivariate analysis multivariate clustering neural networks smoothing spatial point processes statistical distributions statistical tests survival analysis time series analysis

18 Selected methods in Comprehensive R Archive Network (CRAN) Bayesian computation & MCMC, classification & regression trees, genetic algorithms, geostatistical modeling, hidden Markov models, irregular time series, kernel-based machine learning, least-angle & lasso regression, likelihood ratios, map projections, mixture models & modelbased clustering, nonlinear least squares, multidimensional analysis, multimodality test, multivariate time series, multivariate outlier detection, neural networks, non-linear time series analysis, nonparametric multiple comparisons, omnibus tests for normality, orientation data, parallel coordinates plots, partial least squares, periodic autoregression analysis, principal curve fits, projection pursuit, quantile regression, random fields, Random Forest classification, ridge regression, robust regression, Self-Organizing Maps, shape analysis, space-time ecological analysis, spatial analyisis & kriging, spline regressions, tessellations, three-dimensional visualization, wavelet toolbox

19 spatstat: A large CRAN package useful for astronomy ~1500 sta6s6cal func6ons ~800 book describing methods w/ recipes Spa2al Point Pa?erns: Methodology and Applica2ons with R Adrian Baddeley et al D/3D, 1M pts for ploing, 100K exploratory, 10K modeling (on laptop) Smoothing loca6ons and mark variables Global spa6al autocorrela6on func6ons (PCF, K, F, G, J, L*) Tests for inhomogeneity & anisotropy Clustering & hypothesis tests for mul6type data MLE and Bayesian modeling, model residuals & valida6on Poisson, Gibbs, user- supplied models Tessella6ons

20 CRAN Task Views ( CRAN Task Views provide brief overviews of CRAN packages by topic & func6onality. Maintained be expert volunteers. Bayesian ~110 (incl. rbugs, rjags, rstan, MCMCpack, LaplacesDemon, boa) Chem/Phys ~75 packages (incl. 20 for astronomy) Cluster/Mixture ~100 packages Graphics ~40 packages HighPerfComp ~75 packages (parallelia6on, GPUs, etc) Machine Learning ~70 packages Medical imaging ~20 packages Robust ~50 packages Spa6al ~135 packages (incl. spatstat) Survival ~200 packages TimeSeries ~170 packages

21 High Performance Compu2ng with R While originally designed for an individual exploring small datasets, R can be pipelined and can treat megadatsets. Several dozen CRAN packages are devoted to high- performance compu6ng. o foreach func6on released in R by Revolu6on Analy6cs. for loop distributed among local cores o Packages for parallel compu2ng communica2on treat Message Passing Interface, networkspaces, SOCK, PVM, Open MP, o Process management on mul2core computers & clusters: o batch: assists mul6ple parameter simula6ons, permuta6ons, etc o snowfall: wrapper around snow incl. error checks for MPI & PVM o Distributed compu6ng of mul2ple MCMC chains: bugsparallel

22 o Several packages for cloud & grid processing o cloudrmpi: processing using MPI on Amazon s EC2 cloud o GridR: executes R func6ons on remote hosts, clusters or grids o rmr and RHIPE: executes R func6ons via MapReduce on a Hadoop cluster o Several packages treat large memory & out- of- memory data o bigmemory: points to very large objects in memory o ff: fast access to large data on disk o HadoopStreaming: R opera6ons on streaming tabular or string data o Several packages provide GPU compu2ng o gputools: CUDA implementa6on of linear modeling, clustering, SVM, ICA, o magma: matrix algebra for mul6core CPU & GPU architectures o OpenCL: R interface to OpenCL for heterogeneous CPU/GPU systems See CRAN Task View on High Performance Compu)ng hop://cran.r- project.org/web/views/highperformancecompu6ng.html

23 # Support Vector Machine model, predic2on and valida2on Classifica2on in R install.packages('e1071') ; library(e1071) SDSS_svm_mod <- svm(sdss_train[,5] ~.,data=sdss_train[,1:4],cost=100, gamma=1) summary(sdss_svm_mod) ; str(sdss_svm_mod) SDSS_svm_test_pred <- predict(sdss_svm_mod, SDSS_test) plot(sdss_test[,1], SDSS_test[,2], xlim=c(- 0.7,3), col=round(sdss_svm_test_pred), ylim=c(- 0.7,1.8), pch=20, cex=0.5, main='support Vector Machine', xlab='u- g (mag) ) dev.copy2pdf(file='sdss_svm_class.pdf') Support Vector Machine u g (mag) g r (mag) Classification of g r (mag) r i (mag) SDSS Test Dataset r i (mag) i z (mag) www2.astro.psu.edu/users/edf/san6ago_2016

24 Closing remarks Astronomy graduate curriculum needs a year of statistical and computational methodology. Textbook in astroinformatics needed. Some astronomers should get M.S. in statistics or computer science. All astronomers should know methodology at Wikipedia level. Astrostatistics and astroinformatics can become a well-funded, cross-disciplinary research field involving a few percent of astronomers (cf. astrochemists) promulgating modern methodology and pushing its frontiers. Astronomers can benefit from regular use of many methods coded in R & CRAN. R propels implementation of modern procedures with little effort. Experts should investigate utility of R/CRAN in HPC environments.