AI Initiative Seed Grant Awards Announced

Earlier this year, the Office of the Vice President for Research and Dean of the Graduate School announced the University of Maine Artificial Intelligence Initiative. Its vision is to make Maine a world-class hub for AI research, education and applications through its mission to develop transformative AI-based solutions that enhance the social and economic well-being of the citizens of Maine and beyond.

UMaine AI is dedicated to the advancement of AI and its applications — from discovery and learning in foundational AI to its high-impact uses and workforce development. In support of the workforce needs of the state, UMaine offers a multitude of undergraduate and graduate degree programs, interdisciplinary research collaborations and courses that train students in artificial intelligence and its applications. Opportunities abound for undergraduate and graduate students to engage in interdisciplinary research areas, from machine learning, data science and engineering to health and life sciences, business, engineering, and education.

The UMaine AI Initiative is a unique, Maine-based venture that brings together university, industry, government and community collaborators from Maine and beyond to advance the field of artificial intelligence. The initiative leverages the university’s strengths to bring a multi-disciplinary approach to solving problems through cutting-edge research. Additionally, AI related R&D is a major thrust of federal research and funding priorities with a goal to modernize the workforce and maintain US competitiveness in the global economy.

Maine’s land, sea and space grant university addresses grand challenges of global impact and local relevance in AI research projects covering issues from forestry modeling, cancer detection, climate change, space travel and autonomous vehicle programs, to name a few. UMaine AI has four areas of focus: foundational AI research; application of AI; education and workforce development; and social, ethical, policy and legal considerations.

In April, a call for proposals for the UMaine AI Initiative seed grant funding program was released in an effort to help UMaine faculty increase their competitiveness for extramural funding support to advance UMaine’s AI R&D agenda. I am pleased to report that a total of 16 applications were received by the June 1 application deadline and the quality of the proposals from the interdisciplinary teams assembled was outstanding. Although there was not enough funding to advance all of the meritorious proposals that were under consideration, we certainly appreciate the level of interest and range of AI research possibilities that this internal grant program has stimulated. A review committee assembled by the OVPRDGS reviewed, scored, and discussed the applicant pool at length over the past several weeks and recently submitted their funding recommendations to me. Based on those recommendations and an assessment of available resources to dedicate to this strategic initiative, we announce the following four projects will receive seed grant funding and will commence their one year projects on August 1, 2020. Plans are also in place for each of these teams to provide an update on their progress to the university community as part of the UMaine AI initiative’s webinar series.

1. Improved Adversarial Attack Detection Toward Robustness of Deep Neural Networks.

PI: Salimeh Yasaei Sekeh (School of Information and Computer Science (SCIS)). Co-PIs: Ali Abedi (Department of Electrical and Computer Engineering); Richard Corey (VEMI Lab); Nicholas Giudice (VEMI Lab); Collaborator: Theodore Nowak (Pacific Northwest National Laboratory)

Deep neural networks have demonstrated state of the art performance on computer vision and speech recognition tasks in many modern, real-word applications. Despite this success, these models have been shown to be susceptible to adversarial deception and can be fooled by small perturbations to the input referred to as Adversarial Machine Learning (AML). Fooling an autonomous vehicle with adversarial traffic signs or other adversarial objects is a practical, real- world example of such deceptions. Understanding the nature of adversarial machine learning and building robust Machine Learning (ML) models to defend against it, is fundamental to achieving the AI assurance required to apply ML in real-world applications. Hence, as AI research has matured and become more prevalent in commercial systems , characterizing, understanding, and detecting adversarial attacks has increasingly become a topic of interest amongst AI researchers.

2. Advancing Legal-Technological Approaches for Protecting Privacy Rights and Civil Liberties in the Age of Big Data.

PI: Harlan Onsrud (SCIS). Co-PIs: Chaofan Chen (SCIS); Sepideh Ghanavati (SCIS); Manuel Woersdoerfer (SCIS/Maine Business School/Department of Philosophy). Collaborators: Peter Guffen (UMaine Law School/Pierce Atwood LLP) and James Campbell (Maine Freedom of Information Coalition and UMaine adjunct graduate faculty)

Individuals lose control of their identity exposure as artificial intelligence (AI) is incorporated into their everyday lives. This represents a massive societal challenge for the field. Through the application of AI, individuals are influenced almost invisibly in their day-to-day commercial and political decision making to the primary benefit of those using AI methods. In this new environment, substantial loss of personal autonomy has already resulted in many negative societal effects. This research has three core goals. The first is to explore the key privacy and civil liberty principles common across leading AI technology data protection frameworks being advanced around the world. Second, we will postulate the technological information infrastructure that would allow each principle and their potential combinations to be effectively monitored and enforced. This step includes the design of infrastructure components to be demonstrated through a proof of concept prototype that would support an evolving generalizable legal data protection framework. Third, assuming such an infrastructure design, we will identify AI techniques that might be deployed to identify and gather evidence on private and public sector entities that choose to violate the key principles or overall data protection legal framework.

3. Interpretation of Light-Material Interactions with Machine Learning for the Detection of Shared Surface Contamination.

PI: Caitlin Howell (Department of Chemical and Biomedical Engineering). Co-PIs: Salimeh Yasaei Sekeh (SCIS); Sheila Edalatpour (Department of Mechanical Engineering); Richard Corey (VEMI Lab)

The goal of this work is to create a widely accessible and rapid method of assessing the cleanliness of a shared surface. With the spread of the COVID-19 pandemic, our attention to the ability of shared surfaces to transmit disease has become heightened. It has been widely publicized that the SARS-CoV-2 virus can travel in aerosols, landing on surfaces where it can be picked up by others. Furthermore, recent research has demonstrated that infectious copies of the virus can remain viable in the air for up to 3 hours and on surfaces for up to 3 days Kiosks in travel hubs, touch displays at stores, and other shared surfaces have the potential to spread the virus. This issue is not limited to the current outbreak, however. As shared touch surfaces become more and more common, so does their ability to spread any kind of virus, bacterium, or other disease-causing agent that can survive long enough on these surfaces.

4. Context-Dependent Deep Learning for Seabird Recognition in Drone Survey Imagery.

PI: Roy Turner (SCIS). Co-PIs: Cynthia Loftin (U.S. Geological Survey, Maine Cooperative Fish and Wildlife Unit/UMaine Department of Wildlife Ecology); Salimeh Yasaei Sekeh (SCIS).

Image recognition, or more precisely, image segmentation and object recognition, is needed in a vast array of applications, from factory quality control to autonomous automobiles. In natural resources applications, even on this campus, it is used for land cover analysis and tree species detection and marine species identification. Our current project, which was initiated with funds from the Maine Department of Inland Fisheries and Wildlife, US Geological Survey, in-kind support from the US Fish and Wildlife Service, and a 2019 UMS Research Reinvestment Fund (RRF) grant, is acquiring and analyzing diverse types of imagery captured with unmanned aerial vehicles (UAVs) and planes over Maine’s offshore islands and inland rookeries for the purposes of identifying and counting colonial nesting birds with automated technologies. We seek to develop tools to increase efficiency and accuracy of population estimates, an annual responsibility of wildlife management agencies throughout North America, that to date have been based on manual collection and interpretation of collected survey data. Our work directly addresses four areas of University of Maine Artificial intelligence research emphasis: AI in Society, Real Time Machine Learning, Use Inspired AI, and Workforce Development. Our research builds on our partnership across multiple disciplines (Wildlife, Fisheries and Conservation Biology, Forest Resources, Computer Science), engages state and federal collaborators who are providing funding, in-kind support, and professional expertise in our current research this project builds on, and provides mentoring experience for graduate students and real-world learning opportunities for undergraduate students at UMaine Machias and Orono. The connected UMS RRF grant referenced above also involves University of Maine at Machias faculty and students, among other partners.


For questions about the AI Initiative please contact Jason Charland, jason.charland@maine.edu