Michael I. Jordan - US grants

This grant will support research in machine learning for control of nonlinear dynamical systems. Drs. Jordan and Jacobs will study multinetwork connectionist architectures for discovering picewise control strategies, comparing alternative decompositions of control tasks, and learning multiple tasks simultaneously. They expect multinetworks to show faster learning rates, better generalization, more interpretable representations, and more efficient use of hardware than single neural networks.

This Presidential Young Investigator award will enable the investigator to pursue an integrated program of research on the production and perception of speech, including the development of these capacities in children acquiring their native language. The investigator has developed a model of speech production which has become extremely influential in the field. He has identified a number of shortcoming in the model, particularly with respect to questions of the timing of speech. He is now exploring a new family of algorithm which show promise for a better treatment of the timing issues, and which also appear to relate interestingly to phenomena of speech perception. During the ensuing grant period, he will pursue this theoretical development in depth, with particular reference to the problem of modeling inter-work phenomena. The research plans also include further theoretical and empirical work on the problem of speech acquisition. There is a fundamental question of acquisition related to the problem of how it is an infant can learn to perform the complex motor activities of speech based on the acoustic data the child receives as a model. The investigator has developed a hypothesis that speech acquisition is based on the development of an internal model of the vocal tract during babbling, and will be exploring its implications, including implications for the perceptual correlates of coarticulation. The hypothesis makes certain predictions about the order of acquisition of various speech sounds, and another research direction is to test the empirical validity of these predictions.

This award is for a postdoctoral associate in Experimental Science. The associate, David Cohn, will be working on active learning and artificial intelligence.

9404932 Jordan The Associateship in Experimental Science will support a program of interdisciplinary research at the Massachusetts Institute of Technology. The proposed research will focus on probabilistic models for hierarchical neural networks. Two specific architectures will be studied in detail: hierarchical mixtures-of-experts (HME) and Boltzmann trees. HME networks that model mixtures of more complicated probabilistic processes, such as hidden Markov chains and optimal observers will be developed. The expectation-maximization (EM) principle from statistics will be utilized to develop learning algorithms for these architectures. The behavior of these EM algorithms will be analyzed by exploiting the relationship of EM to mean-field theories from statistical physics. Boltzmann machines are a general class of probabilistic model for constraint satisfaction whose development has been hindered by the lack of an efficient learning algorithm. Boltzmann machines with tree-like architectures have simplifying features that make them faster and more efficient. Further work on Boltzmann trees will lead to more powerful algorithms for supervised and unsupervised learning. Various parallels between Boltzmann trees and belief networks for probabilistic reasoning will be explained. ***

This work will exploit learning techniques recently developed to work towards computers that learn to detect and recognize people, estimate user's gestures and communicate visually with them via a photorealistic computer-generated human face. In particular, the plan is to use tow main theoretical and algorithmic approaches to learning: Support Vector Machines and Hidden Markov Models. With these tools, two key aspects of a trainable man-machine interface will be developed: An analysis module that can be trained to estimate in real time facial expressions of the user and associated physical parameters and a synthesis module that can be trained to generate image sequences of a real human face synchronized to a text-to-speech system. The significance of the work is three-fold: (1) The project will contribute to the development of a new generation of computer interfaces more user-friendly and human-centered than today's interfaces. Such interfaces will be of direct use in education and as components of prostheses for the disadvantaged; (2) The project will integrate recently developed learning techniques to real time vision and graphics applications; and (3) The project will explore the boundaries of what is possible to achieve using 2D representations of faces rather than the more common, physically-based, 3D-based models.

Graphical models have become a unifying focus for interdisciplinary research in the areas of probabilistic inference, learning and decision-making. Referred to in various settings as Bayesian networks, Markov random fields, influence diagrams, decision networks, or structured stochastic systems, the graphical model formalism is general enough to encompass a wide variety of classical probabilistic systems in AI and engineering, while providing a firm mathematical foundation on which to design new systems. This research will focus on approximation algorithms for large-scale problems to provide a significantly deeper empirical and theoretical understanding of graphical models. The approach will be based on probability propagation, variational and Monte Carlo methods for inference, learning and decision-making, the aim being to understand the kinds of graphical models for which these methods are appropriate. The PI will extend the scope of approximation methodology to include hybrid graphical models and decision networks, and to provide theoretical convergence analyses and error analyses for them. He will also test out the new methods empirically on standard benchmarks and in a variety of application areas. The ultimate goal of the research is to establish probabilistic graphical models as a full-fledged engineering discipline capable of providing robust, systematic solutions to large-scale problems in inference, learning and decision-making. A successful approximation methodology for graphical models would allow an engineer to design a graphical solution to meet performance specifications for a given problem, where these specifications are given in terms of time / accuracy tradeoffs and estimation / approximation tradeoffs. Even partial progress towards these goals will have wide impact in fields where large-scale probabilistic systems are used, including information retrieval, medical diagnosis, biological sequence analysis, source and error-control coding, speech recognition, and machine vision

The next generation of machine learning systems will need to be substantially more flexible than current systems. Machine learning systems will need to be able to grow new structure as needed, to take into account repeated substructures that arise from relational knowledge, to deal with abstraction hierarchies, and to cope more gracefully with heterogeneous data. This project addresses these issues. It aims at problems both in the Bayesian approach to machine learning (specifically, graphical models) and the frequentist approach to machine learning (specifically, kernel machines). In the graphical model setting, the PI describes a new approach to structure learning based on a flexible prior known as the Chinese restaurant process (CRP)." It explores generalization of the CRP referred to as the hierarchical Dirichlet process" that makes it possible to take into account repeated or partially-repeated sub-structures. It also presents explores a generalization of the CRP that referred to as the nested Chinese restaurant process" for learning abstraction hierarchies.

In the area of kernel machines, the PI builds on his previous NSF-sponsored work to consider methods for combining heterogeneous kernels based on tools from convex optimization, in particular semidefinite programming. He will use these ideas to define novel feature selection methods, and to design new algorithms for the semidefinite programming approach that are the analog of the sequential minimal optimization" (SMO) algorithm for quadratic programming that have permitted the rise to prominence of the support vector machine. The project will focus on driving applications in the areas of information retrieval, bioinformatics, bug-finding in computer programs, and sensor networks.

Pattern classification problems that arise in natural language processing applications, such as parsing, machine translation, and speech recognition, are more complex than those commonly addressed with statistical learning methods. The broad goal of this research project is the design and analysis of statistical learning algorithms that are suitable for these problems. The research is focused on the following questions, which are motivated by characteristic properties of complex pattern classification problems in natural language processing: methods for multiclass classification with desirable statistical and computational properties; methods for structured classification, where the predicted variables come from a large set with a rich structure (for example, predicting the parse tree of a sentence); the extension of these methods to problems with hidden variables, that is, where some relevant data is not observed; and complex nonparametric models for these problems, in particular, computationally efficient nonparametric Bayesian methods based on hierarchical Dirichlet processes. The methods developed will be validated empirically on parsing, machine translation, and speech recognition problems.

The research project is aimed at the development and analysis of statistical learning methods for complex decision problems, such as those that arise in natural language processing. A key goal of research in natural language processing is the development of automated systems, such as translation systems and dialogue systems. The most successful approaches involve the use of statistical methods to exploit language data, such as a text corpus. However, the decision problems that arise are very complex. A good example is the problem of parsing, or recovering the syntactic structure underlying sentences in a language. For such problems, the set of candidate decisions is very large, and possesses considerable structure. This research project is aimed at developing computational and statistical methods that are suitable for complex decision problems of this kind. Successful methods are also likely to have a significant impact in other areas of computer science, including computer vision and bioinformatics, because similar complex decision problems also arise in these areas.

Today's distributed systems are fragile and easily broken. As a result,
total cost of ownership is no longer dominated by capital costs.
Commodity OS's, middleware, and other software building blocks are being
used even to create critical applications such as finance and banking,
yet the complexity of the resulting systems is often beyond our
understanding.

We propose to build on our prior successful argument that system
failures are not problems that will be decisively solved, but ongoing
facts of life to be dealt with. Hence, our approach centers on
systematic fast automatic detection and recovery from many kinds of
failures so fast that failure and recovery will become a form of
adaptation, and we will be able to leverage the well-tested ideas of
control theory, resulting in a new basis for the design of dependable
distributed computing systems. Hence we call our approach RADS: Reliable
Adaptive Distributed Systems. We will develop RADS design guidelines and
prototypes for creating controllable systems by leveraging existing
techniques from Statistical Learning Theory (SLT) and Control Theory
(CT). This will enable much wider applicability of SLT + CT to
dependable computing and establish a concrete venue for collaboration
with those research communities. Although CT and to some degree SLT have
been applied in limited ways for monitoring and optimizing performance,
to our knowledge ours is the first attempt to use these analytical tools
to monitor and control for dependability and high availability.

This proposal develops and studies sparse variants of classic multivariate data analysis methods. It primarily focuses on sparse principal component analysis (PCA) and the related sparse canonical correlation analysis (CCA), but also intends to explore sparse variants of methods such as correspondence analysis and discriminant analysis. The motivation for developing sparse multivariate analysis algorithms is their potential for yielding statistical results that are more interpretable and more robust than classical analyses, while giving up as little as possible in the way of statistical efficiency and expressive power. The investigators have derived a convex relaxation for sparse PCA as a large-scale semidefinite program. The proposed research first studies the theoretical and practical performance of this relaxation as well as the computational complexity involved in solving large-scale instances of the corresponding semidefinite programs. In a next step, it focuses on extending these results to the other multivariate data analysis methods cited above.

Principal Component Analysis (or PCA) is a classic statistical tool used to study experimental data with a very large number of variables (meteorological records, gene expression coefficients, the interest rate curve, social networks, etc). It is primarily used as a dimensionality reduction tool: PCA produces a reduced set of synthetic variables that captures a maximum amount of information on the data. This makes it possible to represent data sets with thousands of variables on a three dimensional graph while still capturing most of the features of the original data, thus making visualization and interpretation easier. Unfortunately, the key shortcoming of PCA is that these new synthetic variables are a weighted sum of all the original variables making their physical interpretation difficult. The proposed research will study algorithms for computing sparse PCA, i.e., computing new synthetic variables that are the weighted sum of only a few problem variables while keeping most of the features of the original data set. Sparse PCA is a hard combinatorial problem but the investigators have produced a relaxation that can be solved efficiently using recent results in convex optimization. The investigators plan to study the theoretical and practical performance of this relaxation and extend these results to other statistical methods.

In this project, the PI and his team will develop a new simulation framework to interactively model and visualize socio-economic and geometric characteristics of urban areas. The framework will consist of a synergistic collaboration of three different areas: behavioral urban modeling, probabilistic graphical modeling, and visualization and computer graphics. In machine learning and statistics, the area of probabilistic graphical modeling offers a flexible framework to build, estimate and simulate from models of substantial complexity and scale, with partially observed data. By accounting for uncertainty and interdependencies, including aspects of dynamic equilibrium that arise in modeling the complex spatio-temporal dynamics of urban areas, the PI argues there is significant potential for breakthroughs in modeling large-scale urban systems. Similarly, by integrating behavioral and geometrical dimensions of urban areas, he expects to exploit the power of behavioral simulations more effectively by filling in geometric details that behavioral models are not well suited to manage, and at the same time provide a powerful framework to generate 2D and 3D geometric representations of urban areas that are behaviorally and geometrically consistent. The PI will take advantage of massive datasets available for urban areas, including parcel and building inventories, business establishment inventories, census data, household surveys, and GIS data on physical and political features, and will fuse these data into a coherent and consistent database to support his modeling objectives. This data fusion will address imputation of missing data, accounting for complex spatial and relational connections among the data sources. The PI will evaluate the accuracy and usability of his system through several deployments in diverse contexts. The PI has elicited engagement from the Urban Land Institute, the European Research Council, and the Council for Scientific and Industrial Research. Several organizations in the San Francisco Bay Area in California and the Puget Sound region in Washington will serve as testbeds for the research. Finally, the PI will collaborate with other NSF-funded research projects, such as the Drought Research Initiative Network, in order to investigate correlations between urban development and water/drought.

Broader Impacts: The results of this multidisciplinary project will have a transformative effect on the area of urban simulation, in that they will enable non-professionals as well as the general public to better understand urban phenomena. City planners, researchers, students, and citizens will be able to efficiently simulate urban processes not previously possible, and to visualize the effects of adopting different urban policies on urban livability and sustainability outcomes, and to address local and global concerns regarding equity, infrastructure, and economic development. The framework will provide interactive desktop and web-based interfaces for configuring urban scenario inputs to a simulation that may reach petabytes in data size, and to visualize the simulation results using 2D aerial views, 3D city walkthroughs, and choroplethic maps and tables of indicators portraying the simulated area. Thus, the work will also advance the fields of visualization and computer graphics, through development of new techniques for large-scale urban modeling and rendering. The PI will develop an open-source system to make the results of this research widely available.

Making Sense at Scale with Algorithms, Machines, and People
University of California, Berkeley

The world is increasingly awash in data. As more and more human activities move on line, and as a growing array of connected devices become integral part of daily life, the amount and diversity of data being generated continues to explode. According to one estimate, more than a Zettabyte (one billion terabytes) of new information was created in 2010 alone, with the rate of new information increasing by roughly 60% annually. This data takes many forms: free-form tweets, text messages, blogs and documents; structured streams produced by computers, sensors and scientific instruments; and media such as images and video.
Buried in this flood of data are the keys to solving huge societal problems, for improving productivity and efficiency, for creating new economic opportunities, and for unlocking new discoveries in medicine, science and the humanities. However, raw data alone is not sufficient; we can only make sense of our world by turning this data into knowledge and insight. This challenge, known as the Big Data problem, cannot be solved by the straightforward application of current data analytics technology due to the sheer volume and diversity of information. Rather, to solve it requires throwing away old preconceptions about data management and breaking down many of the traditional boundaries in and around Computer Science and related disciplines.

The Algorithms, Machines, and People (AMP) expedition at the University of California, Berkeley is addressing this challenge head-on. AMP is a collaboration of researchers with a wide range of data-related expertise, committed to working together to create a new data analytics paradigm. AMP will produce fundamental innovations in and a deep integration of three very different types of computational resources:
1. Algorithms: new machine-learning and analysis methods that can operate at large scale and can give flexible tradeoffs between timeliness, accuracy, and cost.
2. Machines: systems infrastructure that allows programmers to easily harness the power of scalable cloud and cluster computing for making sense of data.
3. People: crowdsourcing human activity and intelligence to create hybrid human/computer solutions to problems not solvable by today's automated data analysis technologies alone.

AMP research will be guided and evaluated through close collaboration with domain experts in key societal applications including: cancer genomics and personalized medicine, large-scale sensing for traffic prediction and environmental monitoring, urban planning, and network security. Advances pioneered by the project will be made widely available through the development of the Berkeley Data Analysis System (BDAS), an open source software platform that seamlessly blends Algorithm, Machine and People resources to solve big data problems.

For more information visit http://amplab.cs.berkeley.edu

? DESCRIPTION (provided by applicant): This proposal responds to the urgent need for advances in data science so that the next generation of scientists has the necessary skills for leveraging the unprecedented and ever-increasing quantity and speed of biomedical information. Big Data hold the promise for achieving new understandings of the mechanisms of health and disease, revolutionizing the biomedical sciences, making the grand challenge of Precision Medicine a reality, and paving the way for more effective policies and interventions at the community and population levels. These breakthroughs require highly trained researchers who are proficient in biomedical Big Data science and have advanced skills at collaborating effectively across traditional disciplinary boundaries. To meet these challenges, UC Berkeley proposes an innovative training program in Biomedical Big Data for advanced Ph.D. students. This training grant will support 6 trainees. We anticipate further extending the reach of our program by admitting up to 2 additional students on alternative support, thus benefitting 8 students per year. The 25 participating faculty have extensive experience with biomedical applications and expertise in biostatistics, causal inference, machine learning, the development of Big Data tools, and scalable computing. Together, they span 8 departments/programs: Biostatistics; Computational Biology; Computer Science; Epidemiology; Integrative Biology; Molecular & Cell Biology; Neuroscience; and Statistics. We will recruit participants from Ph.D. students in their second or third year of study in any/all of these departments. Those accepted into the program will participate in an intensive year of training courses, seminars, and workshops, beginning with introductory seminars in late summer and ending with a capstone project by each participant in the spring. Each trainee will be assigned a secondary thesis advisor with biomedical Big Data science expertise complementing that of the primary thesis advisor. Specialized training will focus on three pillars: (1) translation of biomedical and experimental knowledge and scientific questions of interest into formal, realistic problems of causal and statistical estimation; (2) scalable Big Data computing; and (3) targeted machine learning with causal and statistical inference. Activities will include courses in machine learning targeted learning, statistical programming, and Big Data computing, as well as workshops led by the Berkeley Data Science Institute, Statistical Computing Facility, and Berkeley Research Computing. The capstone course will involve a collaborative project in biomedical science involving the integrated and combined application of skills acquired by the trainees in the three foundational areas. Trainees will also benefit from group seminars, retreats, and interdisciplinary meetings that build a core identity with the cadre and the program. This proposal dovetails with several data science and precision medicine initiatives at UC Berkeley and comes at an ideal time to influence how data science is taught to all graduate students, focusing on biomedical research across campus.

Pressing environmental problems, energy supply security issues, and nuclear power safety concerns drive the worldwide interest in renewable energy. The US Clean Energy Challenge calls for a partnership of states and communities to expand solar to 140GW by 2020. Investment in renewables today is in utility-scale solar plants and wind farms, as well as small-scale distributed rooftop photovoltaics (PV). Large solar plants are cheaper than rooftop PV, but this advantage is diminished when considering transmission infrastructure costs. Generous tax credits and net metering subsidies are responsible for much of the dramatic growth of distributed PV. Under net metering, utilities are mandated to buy back excess generation at retail prices. But tax credits are being phased out, and utilities strenuously oppose net metering policies as they allow PV owners to avoid paying for infrastructure costs and pose an existential threat to utility business models. The growth of distributed PV generation may decelerate. This project aims to sustain and accelerate future growth in distributed PV investments by enabling connected communities to share electricity services. The central thesis is that shared PV ownership and operations can spur greater investment in distributed PV with minimal subsidy, without net metering, and with participants fairly paying for infrastructure, reserves and reliability costs.

Our research will enable connected communities to efficiently use resources, reduce emissions, and support our collective sustainability goals. It will spur deeper penetration of distributed PV without subsidy, while defining new entrepreneurial opportunities in the sharing economy for electricity services. The project will integrate education and research through new interdisciplinary courses that combine technology, economics, policy, and power systems. This research is broadly applicable to other shared services including electricity storage, building energy management, and transportation networks. Specifically, we will (a) develop the infrastructure necessary for sharing electricity services, (b) analyze investment decisions of households under various tariff and subsidy designs, (c) construct behavioral model that predict consumer response to incentives, and (d) conduct an empirical assessment of sharing grounded in data. In our architectural vision for sharing, agents interact with each other through a cloud based supervisory system. This system manages constraints, accepts supply and demand bids for shared resources, clears the market, and publishes prices. A key element of our architecture is software-define-power-flow to scale sharing to millions of clients under a peer-to-peer matching platform. We will make the business case for sharing in the energy sector using game-theoretic methods and micro-economic tools to analyze investment decisions in a sharing economy for electricity services. Recruiting clients to share their resources is a key research challenge. Here, we will apply modern machine learning methods to identify, model, and target suitable clients. Finally, we will use data analytics methods to make a compelling case for sharing based on city-scale data.

In response to NSF's TRIPODS Phase I initiative, the PIs, with expertise in theoretical and applied statistics, computer science, and mathematics at the University of California, Berkeley, will create a Foundations of Data Analysis (FODA) Institute to address cutting-edge foundational issues in interdisciplinary data science. The Institute will advance foundational research and the application of foundational methods through an intensive program of cross-disciplinary outreach to application domains in and beyond the campus research community. In parallel with the massive technological and methodological advances in the underlying disciplines over the past decade, a thriving array of data-related research and training programs has emerged across campus. Yet none of these programs within the campus data science ecosystem are devoted to addressing the interdisciplinary foundations of data analysis in a focused, mission-driven manner. The FODA Institute will address this crucial unmet need. This interdisciplinary project will lay the groundwork for more productive and fruitful interactions between theoretically-inclined data science researchers and researchers in diverse domains that rely upon, but do not always explicitly appreciate, foundational concepts. Advances in this area will lead to more principled extraction of insights from data across a wide range of domains. The three-year Phase I pilot will pave the way for institutionalization of the project as a larger center that will be the subject of a potential Phase II application.

The technical research component of the project addresses four fundamental challenges in data science: the characterization of what is, and what is not, possible in terms of upper and lower bounds for inferential optimization problems; probing more deeply the notion of stability as a computational-inferential principle; exploring the complementary role of randomness as a statistical resource, as an algorithmic resource, and as a tool for data-driven computational mathematics; and developing methods to combine science-based with data-driven models in a principled manner. Each of these challenges addresses old questions in light of new needs, each has important synergies with the other challenges, and each is situated squarely at the interface of theoretical computer science, theoretical statistics, and applied mathematics. The project will bridge the underlying interdisciplinary gaps to address some of the most important questions at the heart of data science today. Funds for the project come from CISE Computing and Communications Foundations and MPS Division of Mathematical Sciences.

This research aims to address the pressing challenges on learning and inference from large-dimensional data. Contemporary sensing and data acquisition technologies produce data at an unprecedented rate. A ubiquitous challenge in modern data applications is thus to efficiently and reliably extract relevant information and associated insights from a deluge of data. In the meantime, this challenge is exacerbated by the unprecedented growth of relevant features one needs to reason about, which oftentimes even outpaces the growth of data samples. Classical statistical inference paradigms, which either only work in the presence of an enormous number of data samples, or ignore the computational cost of the estimators at all, become highly insufficient, or even unreliable, for many emerging applications of machine learning and big-data analytics.

To address the above pressing issues in high dimensions, novel theoretical tools need to be brought in the picture in order to provide a comprehensive understanding of the performance limits of various algorithms and tasks. The goal of this project is four-fold: First, to develop a modern theory to characterize precise performance of classical statistical algorithms in high dimensions. Second, to suggest proper corrections of classical statistical inference procedures to accommodate the sample-starved regime. Third, to develop computationally efficient algorithms that can provably attain the fundamental statistical limits, if possible. Finally, forth, to identify potential computational barriers if the fundamental statistical limits cannot be met. The transformative potential of the proposed research program is in the development of foundational statistical data analytics theory through a novel combination of statistics, approximation theory, statistical physics, mathematical optimization, and information theory, offering scalable statistical inference and learning algorithms. The theory and algorithms developed within this project will have direct impact on various engineering and science applications such as large-scale machine learning, DNA sequencing, genetic disease analysis, and natural language processing. This collaborative program provides cross-university opportunities for students training, and we are committed to engaging and helping underrepresented and women students in STEM through long-term mentorships and outreach activities.This research aims to address the pressing challenges on learning and inference from large-dimensional data. Contemporary sensing and data acquisition technologies produce data at an unprecedented rate. A ubiquitous challenge in modern data applications is thus to efficiently and reliably extract relevant information and associated insights from a deluge of data. In the meantime, this challenge is exacerbated by the unprecedented growth of relevant features one needs to reason about, which oftentimes even outpaces the growth of data samples. Classical statistical inference paradigms, which either only work in the presence of an enormous number of data samples, or ignore the computational cost of the estimators at all, become highly insufficient, or even unreliable, for many emerging applications of machine learning and big-data analytics.

To address the above pressing issues in high dimensions, novel theoretical tools need to be brought in the picture in order to provide a comprehensive understanding of the performance limits of various algorithms and tasks. The goal of this project is four-fold: First, to develop a modern theory to characterize precise performance of classical statistical algorithms in high dimensions. Second, to suggest proper corrections of classical statistical inference procedures to accommodate the sample-starved regime. Third, to develop computationally efficient algorithms that can provably attain the fundamental statistical limits, if possible. Finally, forth, to identify potential computational barriers if the fundamental statistical limits cannot be met. The transformative potential of the proposed research program is in the development of foundational statistical data analytics theory through a novel combination of statistics, approximation theory, statistical physics, mathematical optimization, and information theory, offering scalable statistical inference and learning algorithms. The theory and algorithms developed within this project will have direct impact on various engineering and science applications such as large-scale machine learning, DNA sequencing, genetic disease analysis, and natural language processing. This collaborative program provides cross-university opportunities for students training, and we are committed to engaging and helping underrepresented and women students in STEM through long-term mentorships and outreach activities.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

The Foundations of Data Science Institute (FODSI) brings together a large and diverse team of researchers and educators from UC Berkeley, MIT, Boston University, Bryn Mawr College, Harvard University, Howard University, and Northeastern University, with the aim of advancing the theoretical foundations for the field of data science. Data science has emerged as a central science for the 21st century, a widespread approach to science and technology that exploits the explosion in the availability of data to allow empirical investigations at unprecedented scale and scope. It now plays a central role in diverse domains across all of science, commerce and industry. The development of theoretical foundations for principled approaches to data science is particularly challenging because it requires progress across the full breadth of scientific issues that arise in the rich and complex processes by which data can be used to make decisions. These issues include the specification of the goals of data analysis, the development of models that aim to capture the way in which data may have arisen, the crafting of algorithms that are responsive to the models and goals, an understanding of the impact of misspecifications of these models and goals, an understanding of the effects of interactions, interventions and feedback mechanisms that affect the data and the interpretation of the results, concern about the uncertainty of these results, an understanding of the impact of other decision-makers with competing goals, and concern about the economic, social, and ethical implications of automated data analysis and decision-making. To address these challenges, FODSI brings together experts from many cognate academic disciplines, including computer science, statistics, mathematics, electrical engineering, and economics. Institute research outcomes have strong potential to directly impact the many application domains for data science in industry, commerce, science and society, facilitated by mechanisms that directly involve a stream of institute-trained personnel in industrial partners' projects, and by public activities designed to nurture substantive interactions between foundational and use-inspired research communities in data science. The institute also aims to educate and mentor future leaders in data science, through the further development of a pioneering undergraduate program in data science, and by training a diverse cohort of graduate students and postdocs with an innovative approach that emphasizes strong mentorship, flexibility, and breadth of collaboration opportunities. In addition, the institute plans to host an annual summer school that will deliver core curriculum and a taste of foundational research to a diverse group of advanced undergraduates, graduate students, and postdocs. It aims to broaden participation and increase diversity in the data science workforce, bringing the excitement of data science to under-represented groups at the high school level, and targeting diverse participation in the institute's public activities. And it will act as a nexus for research and education in the foundations of data science, by convening public events, such as summer schools and research workshops and other collaborative research opportunities, and by providing models for education, human resource development, and broadening participation.

The scientific focus of the institute will encompass the full range of issues that arise in data science -- modeling issues, inferential issues, computational issues, and societal issues ? and the challenges that emerge from the conflicts between their competing requirements. Its research agenda is organized around eight themes. Three of these themes focus on key challenges arising from the rich variety of interactions between a decision maker and its environment, not only the classical view of data that is processed in a batch or a stream, but also sequential interactions with feedback (the control perspective), experimental interactions designed to answer "what if" questions (the causality perspective), and strategic interactions involving other actors with conflicting goals (the economic perspective). The other research themes focus on opportunities for major impacts across disciplinary boundaries: on elucidating the algorithmic landscape of statistical problems, and in particular the computational complexity of statistical estimation problems, on sketching, sampling, and sub-linear time algorithms designed to address issues of scalability in data science problems; on exploiting statistical methodology in the service of algorithms; and on using breakthroughs in applied mathematics to address computational and inferential challenges. Intellectual contributions to societal issues in data science will feature throughout this set of themes. The institute will exploit strong connections with its scientific and industrial partners to ensure that these research directions enjoy a rich engagement with a broad range of commercial, technological and scientific application domains. Its sequence of research workshops and a collaborative research program will serve the broader research community by nurturing additional research in these key challenge areas. The institute will be led by a steering committee that will seek the help of an external advisory board to prioritize its research themes and activities throughout its lifetime. Its educational programs will include curriculum development from K-12 through undergraduate, a graduate level visit program, and a postdoc training model, aimed at empowering the next generation of leaders to fluidly work across conventional disciplinary boundaries while being mindful of the full set of scientific issues. The institute will undertake a multi-pronged effort to recruit, engage and support the full range of groups traditionally under-represented in mathematics, computer science and statistics.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.