Schedule for: 23w5094 - Astrostatistics in Canada and Beyond

A buffet dinner is served daily between 5:30pm and 7:30pm in Vistas Dining Room, top floor of the Sally Borden Building.

Let's meet at after dinner to take a walk around town. BIRS Lounge, PDC 2nd floor

Breakfast is served daily between 7 and 9am in the Vistas Dining Room, the top floor of the Sally Borden Building.

A brief introduction to BIRS with important logistical information, technology instruction, and opportunity for participants to ask questions.

A review talk on the history and evaluation of the cross-disciplinary field of Astrostatistics.

A review talk on the big datasets that have been, are begin, and will be generated in astronomy, with a focus on initiatives with Canadian participation.

One-minute lightning talks for participants to introduce themselves and their research.

Lunch is served daily between 11:30am and 1:30pm in the Vistas Dining Room, the top floor of the Sally Borden Building.

Meet in foyer of TCPL to participate in the BIRS group photo. The photograph will be taken outdoors, so dress appropriately for the weather. Please don't be late, or you might not be in the official group photo!

A review talk on some of the statistical learning methods that are useful for Big Data

A review talk on some of the astrostatistics that apply when dealing with datasets involving relatively few data points.

A buffet dinner is served daily between 5:30pm and 7:30pm in Vistas Dining Room, top floor of the Sally Borden Building.

Breakfast is served daily between 7 and 9am in the Vistas Dining Room, the top floor of the Sally Borden Building.

Classification has achieved great success in many applications including speech recognition, handwriting recognition, and biometric identification. Its validity basically relies on the correct labeling of input data. This condition, however, is often violated in many real-world scenarios. As noisy labels severely degrade the performance of many classification algorithms, learning from noisy labels becomes an important task in machine learning. In this paper, we systematically study the problem of binary classification with noisy label and we theoretically investigate the effect of noisy labels. In addition, we identify two scenarios where using the noisy label has the same asymptotic performance as using true label, respectively called balanced mismeasured probabilities and compensatory imbalanced mismeasured. These findings provide important insights into the settings where noise in labels can be ignored. For those cases where using noisy labels incur bias, we propose two correction methods to adjust for the misclassification effects. We theoretically obtain performance bounds for empirical risk minimization in these two correction methods, and further show that the proposed methods have the same asymptotic performance as using the true label. As an application of our work, we consider causal learning where the label of causal-relationship may be mismeasured.

Fast radio bursts (FRBs) are millisecond-duration, bright, extragalactic radio flashes of unknown physical origin. Some FRB sources exhibit repeat bursts, that is, multiple bursts consistent with being emitted from the same physical source. CHIME/FRB (the FRB backend of the Canadian Hydrogen Intensity Experiment) has increased the total number of FRB detections by an order of magnitude with our team's latest catalog release of 536 FRBs. As we run our experiment for longer and our number of FRB detection grows, however, the probability of identifying two or more FRB sources within a typical localization region becomes non-negligible. A question of great importance is then, for a given repeater candidate, what is the probability that each of the bursts in the cluster are physically unrelated to one another (i.e., that they coincided by chance)? In this project, our collaborative research team is working to develop and predict an estimate of the chance coincidence probability of multiple FRBs in the case of a noisy and nonhomogeneous spatial point process.

Neutron stars (NSs) across our vast Universe occasionally merge, unleashing bursts of gravitational waves (GWs) that we can observe here on Earth, starting with their first detection in 2015. Over the next few years, the population of detected mergers will rapidly increase from a few hundred to many million detections per year as new GW observing runs (LIGO/Virgo O4; 2023, LIGO/Virgo O5; 2025) and next-generation detectors (Cosmic Explorer; Einstein Telescope, LISA; 2035) provide data with ever-increasing precision and to larger distances, pushing the reach of gravitational-wave astronomy to the edge of the observable Universe; revolutionizing our view of the cosmos. Making the most of these observations and the rapidly increasing landscape of gravitational-wave detections requires comparing the observed properties, such as their rates, black hole (BH) and NS masses, and BH spins, to theoretical “population synthesis models” simulating their formation pathways. However, at present, this endeavor is limited by the so-called progenitor “Uncertainty Challenge”: uncertainties within the theoretical models are so large, and the models so computationally expensive, that learning about the underlying fundamental physical processes in the lives and deaths of massive stars from observations is completely out of reach, especially for rare events. All present-day simulations thus pay a high price by using highly approximate algorithms that treat the physical processes in a simplified way, or by limiting the total number of simulations, restricting the exploration of the impact of the uncertain physical input assumptions beyond a few variations. In this talk I will introduce the problem and lead an interactive discussion with the participants to investigate statistical techniques to tackle the key Uncertainty Challenge bottleneck in two key areas: (i) improving the sampling of rare events (such as GW sources) in simulations by improving techniques such as adaptive importance sampling, Markov Chain Monte Carlo, and nested sampling and (ii) developing effective emulators predicting model outcomes from small parameter explorations by improving upon techniques from deep learning, normalizing flows, uncertainty quantification, and Gaussian process regression.

An overview of how machine learning methods can be used to speed up and improve standard statistical techniques for fitting models to data.

Neutral patches during reionization leave characteristic damping wing profiles on quasar spectra. If such damping wing profiles could be unambiguously measured, it would pave the way for unraveling the complete reionization history. However, several factors complicate such an endeavor. Notably, a quasar can "erode" its surrounding neutral patches and creates a proximity zone. Therefore, the damping wing only starts from the remaining neutral patches, which could be dozens of comoving Mpc away from the quasar (or ~5000 km/s blueward from the Lya line on the spectrum). Compared with the part redward of the Lya line, spectra features blueward of the Lya line carry potentially more information about the damping wing, yet they are more complex to model due to the extra absorption from the large-scale structure of the universe. The good news is that such large-scale structure could be efficiently sampled from cosmological simulations. In this work, we build a fast simulator to model the quasar absorption spectra and use it to infer the position of the damping wing and the neutral fraction of the universe. We explore different machine learning techniques and find that the simulation-based inference framework can break the degeneracy between parameters and provide accurate measurement of neutral fraction. We tested the method on the mock observation data from with a spectral resolution of 10 km/s and a noise level of 5%, and find that the uncertainty on the size of the quasar HII region is approximately 15% and the uncertainty on neutral fraction is around 0.1 when the universe is >50% neutral.

A framework for data assimilation and prediction of nonlinear dynamics is presented, combining aspects of quantum mechanics, Koopman operator theory, and kernel methods for machine learning. This approach adapts the formalism of quantum dynamics and measurement to perform data assimilation (filtering), using the Koopman operator governing the evolution of observables as an analog of the Heisenberg operator in quantum mechanics, and a quantum mechanical density operator to represent the data assimilation state. The framework is implemented in a fully empirical, data-driven manner by representing the evolution and measurement operators via matrices in a basis learned from time-ordered observations. Applications to data assimilation of the Lorenz 96 multiscale system and others show promising results. Furthermore, our framework provides a route for implementing data assimilation algorithms on quantum computers. This approach can be easily adopted to apply for time-domain astronomy.

Almost all massive galaxies have globular cluster (GC) populations, and the mass of these populations is known to be related to the galaxy’s mass through a linear scaling relation in log10 space. However, many smaller dwarfs do not have any GCs, and these galaxies are often ignored. I have developed a lognormal hurdle model that has the ability to fit a scaling relation to both populations of galaxies simultaneously, and therefore estimate the probability that a galaxy of a given mass has a GC population, as well as the expected mass of that population if it does have one. The model is hierarchical and incorporates measurement errors in both the predictor and response, making it able to fit for galaxy masses and intrinsic scatter as well as the parameters of the scaling relation.

Lunch is served daily between 11:30am and 1:30pm in the Vistas Dining Room, the top floor of the Sally Borden Building.

The topics of these discussions are to be decided mutually and interactively during the meeting. These discussions will occur in conference rooms, with online availability.

The topics of these discussions are to be decided mutually and interactively during the meeting. These discussions will occur in conference rooms, with online availability.

A buffet dinner is served daily between 5:30pm and 7:30pm in Vistas Dining Room, top floor of the Sally Borden Building.

Breakfast is served daily between 7 and 9am in the Vistas Dining Room, the top floor of the Sally Borden Building.

Deep learning models have rapidly proliferated in astronomy — from the Solar System to cosmology — leading to advances of varied import. The black-boxy nature of these algorithms often makes one pause. As we explore the statistical underpinnings of explainability and interpretability in deep learning applications for astronomy, our intent is two-fold. We start by stating the differences between explainability (how models make specific decisions) and interpretability (the broader patterns and rules employed). Using the backdrop of statistics, we will touch upon the significance of interpretability in astronomy and the role of statistical measures in fostering confidence in these tools. Secondly, we hope to foster a dialogue, welcoming insights and perspectives from the attendees to pave the way for a more transparent and accountable future in data-driven astronomical research.

The radial velocity (RV) method is a fruitful approach for detecting exoplanets. Upgraded spectrographs have been designed since the first exoplanet orbiting a Sun-like star was discovered in 1995 using the RV method. The current state-of-the-art spectrographs have the precision and stability necessary to detect low-mass exoplanets such as Earth-analogs. However, the variability in the atmosphere of a star can hide these small exoplanet signals, or even mimic an exoplanet signal leading to false detections. In this presentation, I will introduce the background and highlight the statistical challenges related to this line of research.

A short talk on the the California-Harvard Astrostatistics Collaboration (CHASC) program for multi-dimensional X-ray analysis.

We propose a simple, statistically principled, and theoretically justified method to improve supervised learning when the training set is not representative, a situation known as covariate shift. Building upon a well-established methodology in causal inference, we show that the effect of covariate shift can be reduced or eliminated by conditioning on propensity scores. In practice, this is achieved by fitting learners within strata constructed by partitioning the data based on the estimated propensity scores, leading to approximately balanced covariates and much-improved target prediction. We refer to the overall method as Stratified Learning, or StratLearn. We demonstrate the effectiveness of this general-purpose method on contemporary research questions in cosmology, including the “Supernovae photometric classification challenge”, conditional density estimation of galaxy redshift from photometric data, and redshift calibration for weak lensing. Taken together, these examples illustrate how StratLearn outperforms state-of-the-art importance weighting methods.

A key science case of next-generation radio observatories such as the Square Kilometre Array (SKA) is the direct imaging of the Epoch of Reionization (EoR) using the 21cm line of neutral hydrogen. During the EoR, first-generation galaxies systematically ionize the intergalactic medium (IGM) around them, creating a network of ionized bubbles that grow and eventually merge. Instruments like the SKA are designed to image the Swiss-cheese-like morphology of a bubble-filled IGM. Unfortunately, 21cm data is typically dominated by strong foreground contaminants, and filtering out these foregrounds will mangle the imaged bubbles beyond recognition. In this talk, I will demonstrate how a machine learning approach can leverage non-Gaussianities to allow ionized bubble maps to be recovered from filtered 21cm data. I will also discuss how this relates to high-redshift galaxy surveys, placing a particular emphasis on the combined constraining power of 21cm instruments and telescopes such as JWST.

This short talk discusses the power of using Bayesian methods in stellar evolution.

We present novel statistical methods towards early forecasting of solar flare events, and compare them with machine learning approaches that we have adopted in our previous work. The data sources that we use include: Geostationary Operational Environmental Satellites (GOES), Solar Dynamics Observatory (SDO)/Helioseismic and Magnetic Imager (HMI) and SDO/Atmospheric Imaging Assembly (AIA). The results that I will show in the talk include: (1) strong and weak flare classification with spatial statistics features, together with physics and topological parameters; (2) active region based solar flare intensity prediction with a mixed Long-Short Term Memory regression; and (3) Tensor Gaussian Process with Contraction model for solar flare forecasting combining data of various types and sources.

This short talk will be a very high-level overview of the kinds of open problems in the field of gravitational wave (GW) data analysis to spark discussion. Some problems that be me discussed are: the challenge of analyzing many overlapping signals in LISA data, how to establish the significance of a follow-up detection of a subthreshold GW candidate in LIGO-Virgo data, and/or the successes and limitations of machine learning in GW observations.

The talk will describe the work I have been doing using astronomy image analysis applied to neuroscience images.

Lunch is served daily between 11:30am and 1:30pm in the Vistas Dining Room, the top floor of the Sally Borden Building.

A buffet dinner is served daily between 5:30pm and 7:30pm in Vistas Dining Room, top floor of the Sally Borden Building.

Breakfast is served daily between 7 and 9am in the Vistas Dining Room, the top floor of the Sally Borden Building.

I'll present my newly developed software, AstroPhot, which marks a new state-of-the-art in astronomical image photometry, particularly in uncertainty quantification. As a Python-based, GPU accelerated, object-oriented tool, AstroPhot rivals the capabilities of GALFIT but offers superior speed and adaptability to diverse scientific applications. AstroPhot advances the field by enabling simultaneous multi-band or multi-epoch fitting, which substantially enhances the signal-to-noise ratio. This feature is also present in some conventional photometry codes, along with the capability to concurrently fit overlapping celestial objects such as stars and galaxies. However, AstroPhot goes a step further, introducing the ability to fit the Point Spread Function (PSF) 'live' with other models. As a result, users can extract a comprehensive covariance error matrix for all parameters, including galaxy and PSF parameters. These functionalities pave the way for robust Bayesian analysis projects, notably in rapidly developing PSF-dominated fields such as weak lensing and quasar deblending. They also facilitate any analysis requiring galaxy or stellar parameters derived from Ultraviolet to Infrared photometry. Lastly, I'll introduce a method of utilizing AstroPhot to process multi-epoch data from large surveys such as the Legacy Survey of Space and Time (LSST). This approach promises greater computational efficiency, a more principled technique, and/or superior signal-to-noise extraction than traditional methods like parameter averaging, forced photometry, PSF matching, and simultaneous multi-epoch fitting.

In this talk, I will describe our novel statistical approach and how we use it to understand the formation and evolution of the Milky Way disk.

In this talk, I will discuss statistical methods for the measuring the photometry of faint, crowded, moving point sources.

Ultra-Diffuse Galaxies (UDGs) are a class of extremely faint galaxies that have attracted recent attention in astronomy due to their potential to understand Dark Matter. One important aspect of studying UDGs is their globular clusters (GCs), as many UDGs have been found to possess quite a significant number of GCs despite their low-surface brightness. However, existing GC counting methods in astronomy present various issues. Moreover, there seems to be massive disagreements on the GC counts in the same UDGs. In this paper, we introduce a novel Bayesian model to infer the GC counts in UDGs based on the Mark-Dependently Thinned Point Process (MTPP). Our work is the first substantive application of the MTPP model framework to a complex real-world problem where full Bayesian inference is conducted. We also elucidate some novel insights on the MTPP model framework that provide a holistic view on the nature of the MTPP. Based on MTPP, we treat the point pattern of observed GCs as a realization of a thinned inhomogeneous Poisson process where the thinning depends on both the location and the apparent magnitude of the GC (mark). We construct an efficient adaptive MCMC algorithm to conduct inference for our model. We demonstrate the use of our model on simulated data as well as a set of real GC data obtained from the Program for Imaging of the PERseus (PIPER) survey targeting GCs in the Perseus galaxy cluster.

In the era of radio interferometers and wide-area surveys, the Very Large Array Sky Survey (VLASS) places a strong emphasis on the identification and classification of candidate transient events and hence plays a significant role in mapping the radio sky at high angular resolutions. But, the incomplete imaging of the sky due to the finite number of antennas in VLA along with the snapshot nature of the observations often leads to residual linear imaging artifacts ( linear streaks) seen in Quick-Look (QL) images, which are primarily used for transient searches. While well-established techniques (like CLEAN) maximize the information from snapshot imaging, these techniques can be imperfect. Therefore, knowing whether a part of an image is affected by the above-mentioned linear artifacts is critical for distinguishing truly variable radio sources from artifacts, among other science goals. Moreover, as surveys like VLASS become increasingly common, automatic image-quality classification is increasingly important for rapid data quality assessment and enabling the best science. In this work, we developed a new technique to identify these linear streaks around sources detected in the VLASS Epoch 1 QL images by extending the results of a line detection technique called the Hough Transform. After robustly quantifying the identified streaks, we remove their effects on the sources/components that they overlap. Finally, we use the artifacts-subtracted components’ brightnesses to distinguish them between real astrophysical sources and imaging artifacts.

The luminosity function specifies the distribution of source intensities in a population, providing information regarding the intrinsic characteristics of a class of astronomical objects and their cosmological evolution. In this project, we propose a science-driven hierarchical Bayesian framework to estimate the galaxy luminosity function in X-rays (via a broken power law model, formulated as a mixture of two Pareto distributions), combining non-representative X-ray and optical surveys. The astronomical surveys generally suffer from incompleteness in the low-intensity end, necessitating the application of incompleteness corrections generally estimated from simulations. While this is standard practice for samples selected in a single band, we introduce a methodology for the application of incompleteness corrections for samples selected in two (or more) bands (e.g. X-ray observations of galaxies detected in optical surveys). We account for this incompleteness bias by incorporating an X-ray incompleteness function (estimated from simulations) and an optical incompleteness function (with parameters learned from the observed data) into the model. We avoid detection in the X-ray band for sources detected in the optical band by estimating the source intensities at the locations of the optical sources, using a Poisson model to adjust for background contamination. For the undetected sources in neither band or X-ray sources that are not detected in the optical band we use the corresponding incompleteness functions. The hierarchical model is fit via a Gibbs sampler. We evaluate our model on realistic simulations, mimicking data from the Chandra Deep Field Catalogue, and demonstrate unbiased recovery of the luminosity function even under high proportions of systematic incompleteness. We apply our novel method to data from the Chandra Deep Field South (CDFS).

This talk will provide an overview of simulation-based inference, specific issues with astronomical data, our proposed solutions, and how well this works in practice.

Lunch is served daily between 11:30am and 1:30pm in the Vistas Dining Room, the top floor of the Sally Borden Building.

A buffet dinner is served daily between 5:30pm and 7:30pm in Vistas Dining Room, top floor of the Sally Borden Building.

Breakfast is served daily between 7 and 9am in the Vistas Dining Room, the top floor of the Sally Borden Building.

5-day workshop participants are welcome to use BIRS facilities (TCPL ) until 3 pm on Friday, although participants are still required to checkout of the guest rooms by 11AM.