SIAM News Blog
SIAM News
Print

Data Science at the IMA: An Industry-Supported Initiative

By Benjamin Brubaker, Fadil Santosa, and Daniel Spirn

Industrial Postdoctoral Projects 

    Hossein Keshavarz received his Ph.D. in
   statistics from the University of Michigan
   in 2017. He worked on a natural language
   processing project with Spencer Schaber
   at Cargill. His current Cargill project—in
   collaboration with Hartmut Durchschlag—
   focuses on algorithmic trading strategies
   derived from high-dimensional time series
   analysis, with applications in commodity
   futures. Keshavarz’s research at the IMA
   centers on change point detection in high-
   dimensional, sparse graphical models. In
   this work, he actively collaborates with
   George Michailidis (University of Florida)
   and Yves Atchadé (Boston University).

                    

    
   Stuart Rogers earned his Ph.D. in 2017
   from the University of Alberta, where he
   studied dynamics and control. He worked
   at Lockheed Martin before attending
   graduate school. At Target, Stuart works
   with Kaveh Khodjasteh on recommender
   systems for shipment packaging based on
   mixed-integer linear programming. In
   terms of academia, he studies the
   dynamics and optimal control of rolling ball
   robots and falling cats. Stuart’s
   postdoctoral mentor is Peter Olver
   (University of Minnesota).

                   

   
    Guanglin Xu completed his Ph.D. in
   operations research at the University of
   Iowa in 2017. His IMA mentor is Shuzhong
   Zhang (University of Minnesota). With
   collaborator Andrés Merchán, Xu is
   developing optimal supply chains for
   animal nutrition for Cargill’s feed business.
   His research interest at the IMA is
   optimization under uncertainty, including
   multistage and distributionally-robust
   optimization and their applications in
   healthcare and operations management.

                 

   
    Dongmian Zou received his Ph.D. in 2017
   from the University of Maryland under the
   direction of Radu Balan, who was
   incidentally an IMA industrial postdoc in
   2000. At Cargill, Zou works with Hartmut
   Durchschlag to forecast commodity prices
   and understand factors that drive their
   movement. His research mentor at the IMA
   is Gilad Lerman (University of Minnesota),
   with whom he is using the scattering
   transform to examine fundamental issues
   in graph convolutional neural networks.

                

The Institute for Mathematics and its Applications (IMA) recently established an agreement with Target and Cargill, launching a new era in industrial collaboration at the institute. The two Minnesota-based companies are placing a substantial investment in the mathematical sciences — more than $2 million over a two-year period, likely representing the largest research grant from the industrial sector to a U.S.-based mathematical sciences institute. “The partnership is a win-win for everyone: the IMA, the University of Minnesota (UMN), the two companies, and the region,” Mostafa Kaveh, dean of the College of Science and Engineering at UMN, said. “We are completely behind this initiative and have provided significant institutional support.”

The new initiative fits squarely within the central mission of the IMA: to connect mathematics with its applications. The IMA’s history of collaboration with industry dates back to 1981. Its founders articulated a vision for an interdisciplinary institute with substantial representation from industry and government scientists in the initial proposal submitted to the National Science Foundation (NSF). Since then, the IMA has become the go-to place for collaboration between industry and academia, with the following activities serving as mainstays of the IMA model: (i) industry outreach, (ii) Industrial Problems Seminars, (iii) industrial math modeling workshops, and (iv) industrial postdoctoral fellowships.

After the NSF’s 2015 decision to only partially fund the IMA’s renewal proposal, the institute sought to preserve institutional activities through diversified funding sources. In addition to mitigating the inevitable ups and downs of government funding cycles, this approach aligns the IMA’s incentive structure to better represent its constituents. So the institute devised a new model that makes a compelling case for organizations to support it.

It seemed natural to partner with mathematically sophisticated companies located in the IMA’s own backyard. One of the first companies that the IMA approached was Target. Target’s Chief Data and Analytics Officer, Paritosh Desai—trained in operations research at Stanford University—has a vision for mathematics’ ability to provide companies with a competitive advantage. He was instrumental in shaping the new Data Science Consortium at the IMA and enlisting the participation of Cargill as a second founding member.

Although data science is an inherently interdisciplinary field, its underpinnings are mathematical and algorithmic. “While there are many spectacular successes in machine learning and artificial intelligence, fundamental understanding of their behavior is still quite open,” Desai noted. “This is where we need more mathematics.” Such collaboration provides a meaningful source of interesting and important problems that expand the discipline’s boundaries.

The IMA has previously run programs on the mathematical aspects of data science and is well positioned to address outstanding challenges in the field. Three distinct parts—the industrial postdocs program, the research program, and the training program—reflect the earlier mainstay activities of the industrial portfolio and form the consortium’s core. The industrial postdocs spend half of their time working on directed company research projects. “The three IMA postdoctoral fellows in the Cargill data science group are already making an impact on our business by bringing in new technologies,” James Weed, Vice President of Analytics and Digital Economy at Cargill, said. “We hope that they will consider remaining with Cargill when they are done.” One postdoctoral researcher started at Target last year, and two others will join the company this year. See accompanying sidebar for descriptions of the postdocs’ research projects.

The research program offers topics for thematic semesters, which are determined by consortium members in consultation with the advisory board. The board consists of Desai, Weed, Katherine Ensor (Rice University), Peter Glynn (Stanford), Piotr Indyk (Massachusetts Institute of Technology), Gilad Lerman (UMN), and Joel Tropp (California Institute of Technology). A semester-long program on spatiotemporal forecasting took place in the spring of 2017, and a program on applications of machine learning to supply chains will be held this fall. The training program trains both students and company-based data scientists, providing the companies with both a talent pipeline and cutting-edge tools from academic research. These endeavors enhance the existing core activities at the IMA.

The IMA is using its collaboration with Target and Cargill to lay the foundation for a national data science resource. The initiative has many benefits that extend beyond the partner companies. With its workshops and scientists-in-residence program, the institute continues to function as a center where mathematical scientists can congregate and collaborate. In addition, it acts as a training ground for mathematical scientists at various stages of their careers. The unique industrial postdoc program—admired and imitated around the world—offers industry work experience to recent Ph.D. graduates while retaining connections with academia. It opens career paths in industry while enhancing participants’ academic job prospects. The IMA also sponsors training programs for graduate students interested in transitioning to data science; these programs include an opportunity to work on industry problems — an experience that aids the students’ job search.

IMA leadership has always believed that the most exciting and transformative research arises from problems directly confronting industry, government, and society. This new collaborative program presents yet another avenue to explore these questions and provides exciting possibilities for the future of academic-industrial partnerships in the mathematical sciences.

Benjamin Brubaker is a professor in the School of Mathematics at the University of Minnesota (UMN) and currently serves as the deputy director of the Institute for Mathematics and its Applications (IMA). His research interests lie in analytic number theory and representation theory. Fadil Santosa served as IMA director from 2008 to 2017 and is a professor of mathematics at UMN. He works in inverse problems, optimal design, and photonics. Daniel Spirn is the present director of the IMA, having previously served as deputy director from 2015-2017. He is a professor in the School of Mathematics at UMN and works in analysis of nonlinear partial differential equations.

blog comments powered by Disqus