Inference for MMA



Follow the live streaming in BIDS YouTube channel.

May 29th: (5:30pm) Dinner Reception, Gather Restaurant

Time & Location: May 30th - Berkeley Institute of Data Science (BIDS), Doe Library, UC Berkeley Campus

Wi-Fi: eduroam, AirBears2 (UCB) or CalVisitor (for visitors)

8:30 - 8:40 Logistics, Introduction, Welcome and Summary from Previous Day

Josh Bloom, UC Berkeley Astronomy

The Science of MMA

Moderator, Patrick Brady

8:40 - 8:56 Introduction to MMA

Szabi Marka, Columbia University

8:56 - 9:12 Joint constraints on neutron star equation of state using gravitational waves

Shaon Ghosh, UWM

9:12 - 9:28 Black Hole Binaries and AGN

Zsuzsa Marka, Columbia University

9:28 - 9:44 Numerical Relativity in the Era of Precision Gravity

Dierdre Shoemaker, Georgia Tech

9:44 - 10:00 GW+High-energy neutrinos

Azadeh Keivani, Columbia University

10:00 - 10:16 Gravitational wave parameter and population inference

Richard O’Shaughnessy, Rochester Institute of Technology

10:16 - 10:30 Moderated Discussion

10:30 - 10:40 Break

Inferencing, Emulators & Interpretability

Moderator, David Hogg

10:40 - 10:56 Using Bayesian Inference to search for sub-threshold GW+EM signals

Collin Capano, Albert Einstein Institute, Hannover, Germany

10:56 - 11:12 Deep Learning for Gravitational Wave Astrophysics

Daniel George, Google X

11:12 - 11:28 Spectral Emulators for expensive radiative transfer codes

Wolfgang Kerzendorf, NYU/MSU

11:28 - 11:44 Driving towards interpretability

Ashish Mahabal, Caltech

11:44 - 12:00 Interpretability & Transparency

Daniela Huppenkothen, UW

12:00 - 12:15 Moderated Discussion

12:15 - 1pm Lunch

Frameworks and Approaches to Accelerate Discovery

Moderator, Daniela Huppenkothen

1 - 1:16 An Anomaly Catalog for the Dark Energy Survey

Umaa Rebbapragada, JPL

1:16 - 1:32 Machine and deep learning applications in LSST user generated data products

Federica Bianco, U Delaware

1:32-1:48 ANTARES: Machine Learning on Multi-messenger Alert Streams for Real-time Brokering

Gautham Narayan, STScI

1:48 -2:04 Dimensionality Reduction of SDSS Data with Autoencoders

Stephen Portillo, UW

2:04 - 2:20 Deep learning for the Zwicky Transient Facility: real/bogus classification and identification of fast-moving objects

Dmitry Duev, Caltech

2:20 - 2:35 Moderated Discussion

2:35 - 3:30 Poster Session, Break

Hardware Accelerating Inference and Federated Learning

Moderator, Peter Nugent

3:30 - 3:46 Sharing without Showing: Enabling Secure Collaborative Learning via Cryptography

Wenting Zheng, UC Berkeley

3:46 - 4:02 Probabilistic computing architectures

Eric Jonas, UC Berkeley

4:02 - 4:18 Specialized hardware for machine learning

Amir Khosrowshahi, Intel

4:18 - 4:34 Scalable and Efficient Deep Learning on Supercomputers

Zhao Zhang, Texas Advanced Computing Center

4:34 - 4:50 Moderated Discussion

4:50 - 5:00 Wrap-up and SCiMMA

Patrick Brady, UWM


Federica Bianco, University of Delaware, LSST Science Collaborations

Machine and deep learning applications in LSST user generated data products

Many applications of machine learning and deep learning will be required to generate LSST science beyond the scope of the science and data products that LSST Project has committed to deliver. I will discuss some applications of Deep Learning that the community is planning and scoping, addressing the science cases as well as the potential computational implications.

Collin Capano, Albert Einstein Institute Hannover

Using Bayesian Inference to search for sub-threshold GW+EM signals

The era of multimessenger astronomy with gravitational waves (GW) and electromagnetic (EM) telescopes began with the detection of GW170817. While this event had an exceptionally large signal-to-noise ratio, it hinted at a population of quieter events. By combining data from both GW and EM telescopes, we can discover new, sub-threshold events that would not have been detected otherwise. As an example, I will discuss a potential binary neutron star merger that was identified by combining LIGO data with a simplified model of the Fermi-GBM sky map. New, more confident detections could be made by doing joint Bayesian inference between GW and EM telescopes. Such a framework would further maximize the potential of multimessenger astronomy.

Dmitry Duev, Caltech

Deep learning for the Zwicky Transient Facility: real/bogus classification and identification of fast-moving objects

I will present DeepStreaks and braai, deep-learning systems used by the Zwicky Transient Facility (ZTF) for the identification of fast-moving objects and general real/bogus object classification, respectively. Both systems use convolutional neural networks and demonstrate state-of-the-art performance. I will also report the initials results of the classifier deployment on the Edge Tensor Processing Units (TPUs) that show comparable performance in terms of accuracy, but in a much more (cost-) efficient manner.

Daniel George, Google X

Deep Learning for Gravitational Wave Astrophysics

I will summarize my PhD thesis on applying deep learning using convolutional neural networks for detecting gravitational waves and estimating the parameters of their sources as well as classifying anomalies in LIGO data with state-of-the-art accuracy.

Shaon Ghosh, University of Wisconsin - Milwaukee

Joint constraints on neutron star equation of state using gravitational waves

Gravitational waves emitted from binary systems of compact objects carry with the information of the dynamical evolution of the binary. If the binary system contains neutron star(s), this evolution is strongly influenced by the tidal deformation it(they) suffers, especially in the final few orbits. Inferences on the tidal deformation is therefore possible from the measurement of the emitted gravitational waves. This in turn provides an avenue to constrain the poorly understood equation of state (EoS) of the neutron star. Moreover, the EoS being a fundamental property, we can combine the observations of various events to produce a joint constraint. In this talk I will discuss some of the latest results on constraining of neutrons star EoS using gravitational wave data, and present a new data analysis technique. I will also discuss on how this can be extended to combine data from electromagnetic observations.

Daniela Huppenkothen, University of Washington

Astrophysical Inference in the Era of Machine Learning

Over the past decade, astronomy has adopted a range of new techniques and methods for inferring knowledge from data. Among these, machine learning is quickly becoming a standard toolbox for working with large, complex data sets. While machine learning has certainly made many tasks easier, it also poses new challenges to astronomers: how do we interpret these models? How do we train them? How do we connect the outputs from ML models to the underlying astrophysics we ultimately want to know? How do we quantify biases and uncertainties? In this talk, I aim to set the stage and provide a starting point for discussions around how we use (and plan to use) machine learning in astronomy, and how we can make these models useful for astrophysical inference.

Eric Jonas, UC Berkeley

Probabilistic computing architectures


Azadeh Keivani, Columbia University

Multi-Messenger Searches with Gravitational Waves and High-Energy Neutrinos

Recent discoveries in multi-messenger astrophysics have opened up new windows of exploration to the universe. The detection of gravitational waves (GW) from a binary neutron star merger followed by a short gamma-ray burst recorded the first multi-messenger event involving GWs. In addition, the detection of a high-energy neutrino (HEN) followed by electromagnetic emissions from a flaring blazar was the first compelling evidence for HEN source association. However, no astrophysical source has yet been observed to generate both GWs and HENs. Multi-messenger searches for GWs and HENs provide important insights into the dynamics of and particle acceleration by black holes and neutron stars. The rapid identification of joint signals is crucial for electromagnetic follow-up observations of transient emission that is only detectable for short periods of time. I will present the latest sensitivity of realtime searches of HENs and GWs, and discuss the role of follow-up observations in identifying electromagnetic counterparts to joint GW+HEN events.

Wolfgang Kerzendorf, NYU/MSU

Spectral Emulators for expensive radiative transfer codes

Comparing physical models with observations is one of the main challenges in supernova research. One particularly complex task is the analysis of the evolving spectral sequences. These complex sequences also contain a wealth of information about the object and are thus invaluable to the understanding of these objects. With the profusion of data in the “big data” era, it is essential to have tools that allow automated extraction of physical quantities from the abundance of spectra. We have created a code (TARDIS - Kerzendorf & Sim 2014) that can quickly synthesize supernova spectra with some physical accuracy (using well tested methods). The code is designed to accommodate new physics in a modular form that will allow us to test the systematic uncertainty of several approximations. In addition to the spectral synthesis code, we have created a framework (nicknamed Dalek) that uses machine learning algorithms to find the maximum likelihood of parameters for a given observed spectrum as well as exploring the uncertainties In this talk, I will introduce the code, then will give an overview of some of the preliminary results and will close with an overview of our future research.

Amir Khosrowshahi, Intel

Specialized hardware for machine learning

Compute demand for developing state-of-the-art applications in machine learning and artificial intelligence is increasing at a rapid clip. At the same time, we are hitting perceived limits to the hardware that will be necessary to keep up with this demand. I will discuss the role of hardware in machine learning applications in different domains from systems level and to microarchitecture. I will also present emerging technologies for AI compute such as silicon photonics and room temperature quantum materials.

Ashish Mahabal, Caltech

Driving towards interpretability

Deep learning is everywhere. Its easy to put together a deep learning workflow, and it is fast. The accuracies have also been improving and are generally in the high eighties and nineties. But it is also the PCA of today. It works, but the features that make it work are jumbled and hidden. As we move towards understanding (a) fainter sources, (b) plethora of fleeting sources as required by quick follow-up for LIGO sources, we need to make deep learning more interpretable, be it based on light curves or images. In fact, combining varied sources of knowledge is often the best. We will present some on-going work which combines visualization and a pinch of pragmatism as we handle the growing ZTF data and get ready for even bigger data challenges.

Gautham Narayan, STScI

ANTARES: Machine Learning on Multi-messenger Alert Streams for Real-time Brokering

The Arizona-NOAO Temporal Analysis and Response to Events System (ANTARES) is an alert-broker system applying data science and machine learning techniques to classify objects discovered by the Zwicky Transient Facility (ZTF) in real-time. I will describe our infrastructure, machine learning stages and public interface. In particular, I will detail how we are using deep-learning to enable classification from sparse, early-time data and facilitate rapid follow-up studies. Our value-added data products are available to astrophysicists and can be used to enable new science. We briefly discuss future plans, focusing on scaling from ZTF to the alert volume we expect with the Large Synoptic Survey Telescope (LSST).

Richard O'Shaughnessy, Rochester Institute of Technology

Gravitational wave parameter and population inference

In this talk, I describe a few workflows and algorithms used to interpret of gravitational wave observations of coalescing compact binaries, as a small-scale case study prototyping several challenges for future multimessenger investigations. These strategies statistically distinguish rare transient events from nongaussian noise, infer the distribution of sources consistent with each transient, and characterize the population of astrophysical sources responsible for the data. Focusing on the latter two challenges, I describe some recent progress using improved algorithms, several forms of surrogate modeling, and novel computing architectures and workflows to achieve these goals. I'll then describe new opportunities and challenges associated with rapid population identification and outlier detection; managing and propagating systematic errors in parameter inference; and computing-constrained exploration, approximation, and inference for astrophysical formation scenarios.

Stephen Portillo, University of Washington

Dimensionality Reduction of SDSS Spectra with Autoencoders

We propose using autoencoders to create a data-driven classification of optical galaxy spectra. Autoencoders are unsupervised neural nets that are trained to learn a reduced dimensionality representation of the data. Dimensionality reduction with autoencoders can be thought of as a non-linear generalization of Principal Component Analysis (PCA). We find that, compared to PCA, the autoencoder is better able to reconstruct non-linear features like broad emission lines in QSOs. The autoencoder latent space naturally separates out different classes of galaxies as determined by other criteria like line ratios. Traversing the autoencoder latent space between two spectra produces a series of synthetic spectra that smoothly interpolates between the two. These synthetic spectra look physically plausible with high-level features that vary smoothly.

Umaa Rebbapragada, Jet Propulsion Laboratory

An Anomaly Catalog for the Dark Energy Survey

Identifying unusual or anomalous observations can lead to new discoveries, expose problems in pipeline processing, overturn existing theories, and stimulate new hypotheses. Our objective is to build software tools and methods for the creation of catalogs of astrophysical anomalies. Our process isolates and organizes anomalies in large astronomical survey data sets, and provides explanations to help distinguish between scientifically-interesting astrophysical and unphysical anomalies such as data artifacts and modeling errors. We are prototyping this work on Dark Energy Survey (DES) catalog data, with plans to extend to other survey data. The publication of these types of anomaly catalogs addresses an urgent need to plumb the depths of astronomical archives for new discoveries. The associated software framework provides the community with robust methods for finding, organizing and explaining anomalous observations.

Deirdre Shoemaker, Georgia Tech

Numerical Relativity in the Era of Precision Gravity

Numerical relativity has been a crucial component of the ground-based gravitational wave mission. I'll frame a discussion about what demands will be placed on this field as we look toward future ground and space-based gravitational wave observatories and how we can get the most out of the new era.

Wenting Zheng, UC Berkeley

Sharing without Showing: Enabling Secure Collaborative Learning via Cryptography

Recently, there has been increased interest in collaborative machine learning, where multiple organizations run a training or a prediction task over the joint input data from everyone. Collaboration can be very advantageous because many learning tasks can benefit from learning on complementary datasets or larger datasets. However, when the underlying dataset is sensitive, these organizations cannot collaborate: sensitive data cannot be shared due to factors like privacy concerns, regulatory policies, and/or business competition. A promising paradigm for addressing this problem is provided by secure multi-party computation (MPC), a classic cryptographic technique that allows multiple parties to compute a generic function on everyone’s inputs without revealing the inputs. At a high level, MPC runs a computation on encrypted inputs and produces an encrypted final result that the participants can then jointly decrypt. In this talk, I will give an overview of some current and exciting research in secure collaborative learning using MPC.

Zhao Zhang, Texas Advanced Computing Center

Scalable and Efficient Deep Learning on Supercomputers

Scientists across fields are exploring and exploiting DL techniques for classification, prediction, and simulation dimensionality reduction. These DL applications are naturally supercomputing applications given the computation, communication, and I/O characteristics. In this talk, I will present two works to enable highly scalable distributed DL training. The first one is to enable efficient and scalable l/O for DL applications on supercomputers with FanStore, with which we are able to scale real world applications to hundreds nodes on CPU and GPU cluster with over 90% scaling efficiency. The second one focuses on scaling practice and its application in ImageNet training on thousands of compute nodes with the state-of-the-art validation accuracy.

Poster's Abstracts