Please join us for the MINDS & CIS Seminar Series
Tuesday, June 2, 2020 at 12:00 pm Eastern Time (US and Canada)
“Identification Theory in Segregated Graph Causal Models” by Eli Sherman (CS, JHU)
Seminar will be remote via Zoom
Join Zoom Meeting
https://wse.zoom.us/j/95178918871?pwd=bjh0M0VjMHczVXdZMnlxN0xyR2VqUT09
Meeting ID: 915 7891 8871
Password: clark_hall
Abstract – In recent years there has been an explosion of interest in causal inference methodologies in the machine learning and broader data science communities. A key issue at the foundation of all causal analysis is the concept of identification: the work of Pearl and others has sought to formally characterize when causal queries are estimable from available (i.e. observed) data given the assumed causal model. In this talk I’ll discuss extensions to the classical latent-variable DAG identification framework that are particularly relevant when there is dependence among data samples, such as social networking and spatial settings. Toward this end, I will introduce the segregated graph model, a super model for latent-variable DAGs, and argue for its use in these dependent settings. I will then provide sound and complete identification results for ‘node’ (i.e. classical) interventions. Finally, I will describe sound and complete results for the identification of ‘policy’ interventions, corresponding to a sequential decision-making setting, in segregated graphs and demonstrate how these results generalize and nest several existing identification results.
Bio – Eli Sherman is a PhD student in the Computer Science Department at Johns Hopkins University. He develops methods for obtaining causal inferences in social networking and dependent data contexts as well as approaches for intervention tailoring. He is interested in applications of these methods to healthcare, economics, and public policy. Eli is supervised by Ilya Shpitser and is affiliated with the Malone Center for Engineering in Healthcare and the Mathematical Institute for Data Science, from which he receives support through the MINDS PhD Fellowship.
Machine Learning: Dynamical, Statistical, and Economic Perspectives
Abstract: Much of the recent focus in machine learning has been on the pattern-recognition side of the field. I will focus instead on the decision-making side, where many fundamental challenges remain. Some are statistical in nature, including the challenges associated with multiple decision-making. Others are economic, involving learning systems that must cope with scarcity and competition. I’ll pose, and perhaps even solve, a few algorithmic problems in these areas, making use of a line of recent work on (continuous-time) dynamical systems perspectives on optimization and diffusion.
Michael I. Jordan, University of California, Berkeley, U.S.
This is one of seven virtual plenary talks originally scheduled for the 2020 SIAM Conference on Mathematics of Data Science. For more information on this session, visit https://meetings.siam.org/sess/dsp_programsess.cfm?SESSIONCODE=69238. To view the virtual program and register for other invited plenary talks, minitutorial talks, and minisymposia, please visit the MDS20 website at https://www.siam.org/conferences/cm/conference/mds20.
Graphical Models in Machine Learning, Networks, and Uncertainty Quantification
Abstract: This talk is an overview of recent work graph models for classification using similarity graphs, for community detection in networks, and for the subgraph isomorphism problem in multichannel networks. The equivalence between the graph mincut problem and total variation minimization on the graph allows one to cast graph-cut variational problems in the language of total variation minimization, thus creating a parallel between low dimensional data science problems in Euclidean space (e.g. image segmentation) and high dimensional clustering. Semi-supervised learning with a small amount of training data can be carried out in this framework with diverse applications ranging from hyperspectral pixel classification to identifying activity in police body worn video. It can also be extended to the context of uncertainty quantification with Gaussian noise models. The problem of community detection in networks also has a graph-cut structure and algorithms are presented for the use of threshold dynamics for modularity optimization. With efficient methods, this allows for the use of network modularity for unsupervised machine learning problems with unknown number of classes. Finally we discuss a different class of graph problem – namely identifying template structure in large world graphs and how combinatorial filtering methods can be structured to efficiently attack this problem.
Andrea L. Bertozzi, University of California, Los Angeles, U.S.
This is one of seven virtual plenary talks originally scheduled for the 2020 SIAM Conference on Mathematics of Data Science. For more information on this session, visit https://meetings.siam.org/sess/dsp_programsess.cfm?SESSIONCODE=69236. To view the virtual program and register for other invited plenary talks, minitutorial talks, and minisymposia, please visit the MDS20 website at https://www.siam.org/conferences/cm/conference/mds20.
Facts, Flexibility, and Fairness
Abstract: The vast majority of work in algorithmic fairness addresses classification and scoring tasks. Missing from the literature is a treatment of fairness in ranking, a central element in many selection procedures as well as in approaches to affirmative action. Definitions of (algorithmic) fairness fall roughly into two categories: group fairness and individual fairness. The former typically require that some statistic be similar across supposedly disjoint groups; the latter require that people who are similar with respect to a given task should be treated similarly. A more recent line of work lies somewhere in between, and considers a large number of possibly intersecting subpopulations. The philosophy that undergirds this last notion is to first try to understand what the historical outcomes data say (“facts”); next explore the evidence-consistent options (“flexibility”); and, finally, adjust as appropriate (“fairness”). This self-contained talk will focus on evidence-based ranking and the flexibility it offers, noting that flexibility can be a double-edged sword.
Cynthia Dwork, Harvard University, U.S.
This is one of seven virtual plenary talks originally scheduled for the 2020 SIAM Conference on Mathematics of Data Science. For more information on this session, visit https://meetings.siam.org/sess/dsp_programsess.cfm?SESSIONCODE=69237. To view the virtual program and register for other invited plenary talks, minitutorial talks, and minisymposia, please visit the MDS20 website at https://www.siam.org/conferences/cm/conference/mds20.
Graphons and Machine Learning: Modeling and Estimation of Sparse Networks at Scale
Abstract: There are numerous examples of sparse network at scale, for example the Internet, the WWW, online social networks, and large bipartite networks used for recommendations. How do we model and learn these networks? In the case of relatively small or moderately sized networks, it’s appropriate to model the network parametrically, and attempt to learn these parameters. For networks at scale, a non-parametric representation is more appropriate. In this talk, I first review the theory of graphons, developed over the last 15 years to describe limits of dense graphs, and the more the recent theory describing sparse graphs of unbounded average degree, including power-law graphs. I then show how to use these graphons as non-parametric models for sparse networks. I show how to get consistent estimators of these non-parametric models, and moreover how to do this in a way that protects the privacy of individuals on the network. Finally, I provide algorithms based on these models, for example for recommendation algorithms in sparse bipartite networks such as Netflix.
Jennifer Tour Chayes, Microsoft Research, U.S.
This is one of seven virtual plenary talks originally scheduled for the 2020 SIAM Conference on Mathematics of Data Science. For more information on this session, visit https://meetings.siam.org/sess/dsp_programsess.cfm?SESSIONCODE=69232. To view the virtual program and register for other invited plenary talks, minitutorial talks, and minisymposia, please visit the MDS20 website at https://www.siam.org/conferences/cm/conference/mds20.
The Work of Theory in the Age of Computational Supremacy
Abstract: The last decade has produced an expansion in computational power that far outdistances the rate of expansion in previous decades. Driving forces include the global spread of smartphones and the epic buildout of the associated communications and computation infrastructure; as a trivial corollary this infrastructure makes available cloud computing resources able to effortlessly crush practical challenges that were previously considered unimaginably difficult. A deeper corollary has been the rapid transition to dominance of empirical research based almost exclusively on massive computational trial and error experimentation. The progress over the last decade in deep learning for image understanding and for natural language processing is a famous example -— and maybe the first ever — of stunning progress being driven by such computational trial and error. The recent transition to effortless access to massive computational capabilities is a one-time transition in human history. However, not all of us — especially not all of us theorists — recognize that this transition has occurred. How is this new era different for theorists? What attitudes and ideas need to change? I will give two case studies illustrating these points, one based on work of Stephane Mallat and co-authors developing competitors to deep nets based on harmonic analysis and one based on work of Vardan Papyan on about spectral features perceptible in the spectra of deepnet Hessians.
David L. Donoho, Stanford University, U.S.
One of seven virtual plenary talks originally scheduled for the 2020 SIAM Conference on Mathematics of Data Science. For more information on this session, visit https://meetings.siam.org/sess/dsp_programsess.cfm?SESSIONCODE=69238. Please visit the MDS20 website at https://www.siam.org/conferences/cm/conference/mds20.
The Deep Learning – Applied Math Connection
Abstract: Deep learning (DL) is causing revolutions in computer perception, signal restoration/reconstruction, signal synthesis, natural language understanding, and process control. DL is increasingly used to provide approximate solutions to PDE and non-linear optimization problems, with many applications in cosmology, material science, high-energy physics, and various applications of fluid dynamics. But one of the most direct impacts of DL on the scientific computing community has been to provide flexible software platforms for numerical problems, such as PyTorch and TensorFlow, with built-in support for multi-dimensional arrays, GPU, parallelism, and automatic differentiation.
While DL has become a new tool in the toolbox of the applied mathematician, DL heavily relies on the tools of applied mathematics, such as large-scale non-linear, non-convex optimization. But our understanding of the landscape of the objective function and the convergence properties in DL systems is still very superficial.
One of the key questions in Machine Learning today is how to learn predictive models of the world in a self-supervised manner, a bit like humans and animals. A class of methods that can predict high-dimensional signals, such as video, under uncertainty will be presented. It is based on capturing dependencies between variable by shaping an energy function.
Yann LeCun, Facebook, U.S.
This is one of seven virtual plenary talks originally scheduled for the 2020 SIAM Conference on Mathematics of Data Science. For more information on this session, visit https://meetings.siam.org/sess/dsp_programsess.cfm?SESSIONCODE=69233. To view the virtual program and register for other invited plenary talks, minitutorial talks, and minisymposia, please visit the MDS20 website at https://www.siam.org/conferences/cm/conference/mds20.
Inexact Accelerated High-order Proximal-point Methods
Abstract: In this talk, we present a new framework of Bi-Level Unconstrained Minimization (BLUM) for the development of accelerated methods in Convex Programming. These methods use approximations of the high-order proximal points, which are solutions of some auxiliary parametric optimization problems. For computing these points, we can use different methods, and, in particular, the lower-order schemes. This opens a possibility for the latter methods to overpass the traditional limits of the Complexity Theory. As an example, we obtain a new second-order method with the convergence rate O(k^(−4)), where k is the iteration counter. This rate is better than the maximal possible rate of convergence for this type of methods, as applied to functions with Lipschitz continuous Hessian. We also present new methods with the exact auxiliary search procedure, which have the rate of convergence O(k^(−(3p+1)/2)), where p≥1 is the order of the proximal operator. The auxiliary problem at each iteration of these schemes is convex.
Yurii Nesterov, Université Catholique de Louvain, Belgium
This is one of seven virtual plenary talks originally scheduled for the 2020 SIAM Conference on Mathematics of Data Science. For more information on this session, visit https://meetings.siam.org/sess/dsp_programsess.cfm?SESSIONCODE=69235. To view the virtual program and register for other invited plenary talks, minitutorial talks, and minisymposia, please visit the MDS20 website at https://www.siam.org/conferences/cm/conference/mds20.