Last Updated: 20171211

AAS 231: Special Sessions #225 & #239

Applied Statistical Methods in Astronomy:
Gaussian Processes and Machine Learning

Wednesday, January 10, 2018

2:00pm-3:30pm // National Harbor 2
| Description | Schedule | Contacts | Posters | changelog |


Modern telescopes provide challenging data, not only in quantity but also quality, demanding new methods and techniques for scientific inference. New algorithms specific to astronomical problems are being developed and brought to the community by a new generation of scientists. This special session will focus on statistical methods that were published within the last year. This session originates from the AAS Working Group on Astroinformatics and Astrostatistics and the CHASC International Center for Astrostatistics. The goal of this Special Session is to review advances in the newly popular methods of gaussian processes and machine learning, to present applications to data, and to discuss current issues and future perspectives. These methods have applications across the entire spectrum of astronomical research and are being rapidly developed. They have to be presented at large forums to make the community aware of the rapid progress being made in this field. The specific invited talks include discussion of application of gaussian processes to time-series spectra in exoplanet research, machine learning techniques to quantify variability states of a micro-quasar, machine learning application in cosmology, and source detection. We also intend to have an associated poster session allowing contributions from the entire community. The session schedule will allow for a discussion and input from the audience.


Chair: Aneta Siemiginowska (CfA/ CHASC)

2:02 pm: Dan Foreman-Mackey (Center for Computational Astrophysics, Flatiron Institute)

An astronomer's introduction to Gaussian Processes

Abstract: Gaussian Processes (GPs) are a class of stochastic models that have become widely used in astronomy. A general introduction to GP modeling can be mystifying so, in this talk, I will introduce GP modeling with a focus on applications from the astronomy literature. I will summarize the basic theory and motivate the broad applicability of these methods. Finally, I will discuss some of the computational limitations of simple implementations of GP models and describe some recent developments that make these models more broadly tractable.

Presentation slides [.pdf]
Dan Foreman-Mackey is an Associate Research Scientist at the Flatiron Institute's Center for Computational Astrophysics.

2:22 pm: Ian Czekala (Stanford)

Using Gaussian Processes to Construct Flexible Models of Stellar Spectra

Abstract:The use of spectra is fundamental to astrophysical fields ranging from exoplanets to stars to galaxies. In spite of this ubiquity, or perhaps because of it, there are a plethora of use cases that do not yet have physics-based forward models that can fit high signal-to-noise data to within the observational noise. These inadequacies result in subtle but systematic residuals not captured by any model, which complicates and biases parameter inference. Fortunately, the now-prevalent collection and archiving of large spectral datasets also provides an opening for empirical, data-driven approaches. We introduce one example of a time-series dataset of high-resolution stellar spectra, as is commonly delivered by planet-search radial velocity instruments like TRES, HIRES, and HARPS. Measurements of radial velocity variations of stars and their companions are essential for stellar and exoplanetary study; these measurements provide access to the fundamental physical properties that dictate all phases of stellar evolution and facilitate the quantitative study of planetary systems. In observations of a (spatially unresolved) spectroscopic binary star, one only ever records the composite sum of the spectra from the primary and secondary stars, complicating photospheric analysis of each individual star. Our technique "disentangles" the composite spectra by treating each underlying stellar spectrum as a Gaussian process, whose posterior predictive distribution is inferred simultaneously with the orbital parameters. To demonstrate the potential of this technique, we deploy it on red-optical time-series spectra of the mid-M-dwarf eclipsing binary LP661-13, which was recently discovered by the MEarth project. We successfully reconstruct the primary and secondary stellar spectra and report orbital parameters with improved precision compared to traditional radial velocity analysis techniques.

Presentation slides [.pdf]
Ian Czekala is a KIPAC Postdoctoral Fellow at Stanford University.

2:38 pm: Daniela Huppenkothen (University of Washington)

Classifying Black Hole States with Machine Learning

Abstract:Galactic black hole binaries are known to go through different states with apparent signatures in both X-ray light curves and spectra, leading to important implications for accretion physics as well as our knowledge of General Relativity. Existing frameworks of classification are usually based on human interpretation of low-dimensional representations of the data, and generally only apply to fairly small data sets. Machine learning, in contrast, allows for rapid classification of large, high-dimensional data sets. In this talk, I will report on advances made in classification of states observed in Black Hole X-ray Binaries, focusing on the two sources GRS 1915+105 and Cygnus X-1, and show both the successes and limitations of using machine learning to derive physical constraints on these systems.

Presentation slides [.pdf]
Daniela Huppenkothen is an Associate Director for the DIRAC Institute at the University of Washington, Seattle.

2:54 pm: Michelle Ntampaka (Harvard University)

Dynamical Mass Measurements of Contaminated Galaxy Clusters Using Support Distribution Machines

Abstract: We study dynamical mass measurements of galaxy clusters contaminated by interlopers and show that a modern machine learning (ML) algorithm can predict masses by better than a factor of two compared to a standard scaling relation approach. We create two mock catalogs from Multidark~s publicly available N- body MDPL1 simulation, one with perfect galaxy cluster membership infor- mation and the other where a simple cylindrical cut around the cluster center allows interlopers to contaminate the clusters. In the standard approach, we use a power-law scaling relation to infer cluster mass from galaxy line-of-sight (LOS) velocity dispersion. Assuming perfect membership knowledge, this unrealistic case produces a wide fractional mass error distribution, with a width E=0.87. Interlopers introduce additional scatter, significantly widening the error distribution further (E=2.13). We employ the support distribution machine (SDM) class of algorithms to learn from distributions of data to predict single values. Applied to distributions of galaxy observables such as LOS velocity and projected distance from the cluster center, SDM yields better than a factor-of-two improvement (E=0.67) for the contaminated case. Remarkably, SDM applied to contaminated clusters is better able to recover masses than even the scaling relation approach applied to uncon- taminated clusters. We show that the SDM method more accurately reproduces the cluster mass function, making it a valuable tool for employing cluster observations to evaluate cosmological models.

Presentation slides [.pdf]
Michelle Ntampaka is a Data Science Initiative Fellow at Harvard University.

3:10 pm: Tamas Budavari (Johns Hopkins University)

Sub-Band Image Reconstruction using DCR

Abstract:Refraction by the atmosphere causes the astrometric positions of sources to depend on the airmass through which an observation was taken. This shift is dependent on the underlying spectral energy of the source and the filter or bandpass through which it is observed. Wavelength- dependent refraction within a single passband is often referred to as differential chromatic refraction (DCR). With a new generation of astronomical surveys undertaking repeated observations of the same part of the sky over a range of different airmasses and parallactic angles DCR should be a detectable and measurable astrometric signal. Here we introduce a novel procedure that uses this subtle signal to infer the underlying spectral energy distribution of a source; we solve for multiple latent images at specific wavelengths via a generalized deconvolution procedure built on robust statistics.

Presentation slides [.pdf]
Tamas Budavari is an Assistant Professor in Applied Math at the Johns Hopkins University.

Poster Session 239

A list of poster presentations accepted for the session:
239.01 - Recovering Galaxy Properties Using Gaussian Process SED Fitting -- Kartheik Iyer (Rutgers)
239.02 - SESNPCA: Principal Component Analysis Applied to Stripped-Envelope Core-Collapse Supernovae -- Marc Williamson (NYU)
239.03 - Gravitational Lensing Mass Mapping with Gaussian Processes -- Michael Schneider (Livermore)
239.04 - Analyzing LSST-like Quasar Light Curves with Kriging -- Jared Haughton (Drexel University)
239.05 - Learning the Relationship between Galaxy Spectra and Star Formation Histories -- Christopher Lovell (Sussex)


Vinay Kashyap (vkashyap @ cfa . harvard . edu)
Aneta Siemiginowska (asiemiginowska @ cfa . harvard . edu)


2017-dec-12: set up page.


Applied Statistical Methods in Astronomy
Gaussian Processes
Machine Learning

Wed 10 Jan 2018
National Harbor 2

Dan Foreman-Mackey
Ian Czekala
Daniela Huppenkothen
Michelle Ntampaka
Tamas Budavari