Speakers
Deep Learning Data-driven approaches for Epidemic Forecasting
The devastating impact of the currently unfolding global COVID-19 pandemic and those of the Zika, SARS, MERS, and Ebola outbreaks over the past decade has sharply illustrated our enormous vulnerability to emerging infectious diseases. There are many questions that are being studied by epidemiologists and public officials during these outbreaks. Building on our prior work, we have been pursuing multiple activities amidst the COVID-19 pandemic in the United States, collaborating with partners in academia, industry and public health agencies, from award-winning work on helping forecast pandemic trajectories (also shown on the CDC website) to designing more localized and less burdensome campus interventions. In this talk, I will briefly give an overview of our recent research in designing well calibrated, robust, accurate and interpretable deep learning models for epidemic forecasting, illustrating the important role data science and machine learning have to play for pandemic prevention and prediction.
Bio
B. Aditya Prakash is an Associate Professor in the College of Computing at the Georgia Institute of Technology (“Georgia Tech”). He received a Ph.D. from the Computer Science Department at Carnegie Mellon University
in 2012, and a B.Tech (in CS) from the Indian Institute of Technology (IIT) -- Bombay in 2007. He has published one book, more than 80 papers in major venues, holds two U.S. patents and has given several tutorials at leading conferences.
His work has also received multiple best-of-conference, best paper and travel awards. His research interests include Data Science, Machine Learning and AI, with emphasis on big-data problems in large real-world networks and time-series,
with applications to computational epidemiology/public health, urban computing, security and the Web. Tools developed by his group have been in use in many places including ORNL, Walmart and Facebook. He has received several awards
such as a Facebook Faculty Award, the NSF CAREER award and was named as one of ‘AI Ten to Watch’ by IEEE. His work has also won awards in multiple data science challenges (e.g the Facebook COVID19 Symptom Challenge) and been highlighted
by several media outlets/popular press like FiveThirtyEight.com. He is also a member of the infectious diseases modeling MIDAS network and core-faculty at the Center for Machine Learning (ML@GT) and the Institute for Data Engineering
and Science (IDEaS) at Georgia Tech. Aditya’s Twitter handle is @badityap.
Modeling raw, messy time series data with latent stochastic differential equations
Much real-world data is sampled at irregular intervals, but most time series models require regularly-sampled data. Continuous-time models address this problem, but until now only deterministic models (based on ordinary differential equations) or linear-Gaussian
models were efficiently trainable with millions of parameters. We construct a scalable algorithm for computing gradients through samples from stochastic differential equations (SDEs), and for gradient-based stochastic variational
inference in function space, all with the use of adaptive black-box SDE solvers. This allows us to fit a new family of richly-parameterized distributions over time series, in which neural networks can parameterize both dynamics
and likelihoods. We demonstrate these latent SDEs on motion capture data, and provide an open-source PyTorch library for fitting large SDE models.
[Slides]
[Paper]
[Code]
Bio
David Duvenaud is an assistant professor at the University of Toronto. His research focuses on constructing deep probabilistic models to help predict, explain and design things. For example: Neural ODEs, a kind of
continuous-depth neural network, Automatic chemical design using generative models, Gradient-based hyperparameter tuning, Structured latent-variable models for modeling video, and Convolutional networks on graphs. Previously, He
was a postdoc in the Harvard Intelligent Probabilistic Systems group with Ryan Adams. He did my Ph.D. at the University of Cambridge, where his advisors were Carl Rasmussen and Zoubin Ghahramani. His M.Sc. advisor was Kevin Murphy
at the University of British Columbia. He spent a summer working on probabilistic numerics at the Max Planck Institute for Intelligent Systems, and the two summers before that at Google Research, doing machine vision. He co-founded
Invenia, an energy forecasting and trading firm where he still consult. He is also a founding member of the Vector Institute.
Irrational Exuberance: Why we should not believe 99% of papers on Time Series Anomaly Detection
In the last five years there has been an explosion of paper on time series anomaly detection (TSAD) appearing in all the top conferences. In this talk I will make a surprising claim. Almost all such papers suffer from various flaws, including: Testing
on deeply flawed datasets, use of inappropriate measures of success, non-reproducible experiments, unjustified complexity, and ignoring competitive decades-old methods. I will demonstrate that because of these flaws, we should
not believe the claims of most papers on time series anomaly detection. Rather than a completely pessimistic talk, I will take this opportunity to release a new set of 250 benchmark datasets and guidelines that will go some way
to mitigating the problems.
[Slides]
Bio
Eamonn Keogh is a Distinguished Professor of Computer Science at the University of California Riverside. With his students, he invented many of the most commonly used time series primitives, including Shapelets, Discords,
Motifs, SAX, PAA, LB_keogh and the Matrix Profile. He has won at least one best paper award in the major data mining conferences (SIGMOD, SIGKDD, ICDM, SDM) and his h-index of 100 reflects the significant amount of citations his
work has attracted.
On the Interface between Optimal Transport and Time Series
I will start this talk with a short intro to optimal transport theory, and mention a few interesting areas at the intersection of both optimal transport and time series. I will describe in more detail the JKO proximal gradient descent scheme and show how it can be used to model time series of snapshots of populations (joint work with Charlotte Bunne, Laetitia Papaxanthos and Andreas Krause) and also the Wasserstein DTW discrepancy for such time series.
Bio
Marco Cuturi joined Google Brain, in Paris, in October 2018. He graduated from ENSAE (2001), ENS Cachan (Master MVA, 2002) and holds a PhD in applied maths obtained in 2005 at Ecole des Mines de Paris. He worked as a
post-doctoral researcher at the Institute of Statistical Mathematics, Tokyo, between 11/2005 and 03/2007. He worked in the financial industry between 04/2007 and 09/2008. After working at the ORFE department of Princeton University
between 02/2009 and 08/2010 as a lecturer, he was at the Graduate School of Informatics of Kyoto University between 09/2010 and 09/2016 as a tenured associate professor. He then joined ENSAE, the french national school for statistics
and economics, in 9/2016, where he still teaches. His recent proposal to solve optimal transport using an entropic regularization has re-ignited interest in optimal transport and Wasserstein distances in the machine learning community.
His work has recently focused on applying that loss function to problems involving probability distributions, e.g. topic models / dictionary learning for text and images, parametric inference for generative models, regression with
a Wasserstein loss and probabilistic embeddings for words.
