I was always motivated to do research that could be of service to my community. I am deeply interested in how developments in the technology we build get translated into things that are accessible and can be used by people beyond our field of expertise.
I'm a professor in electrical engineering and computer science, and specifically in the AI and decision-making sub department at MIT. I'm also a member of CSAlL (Computer Science and AI Laboratory). My research group is fundamentally interdisciplinary. We focus on the development of novel AI, computer vision, machine learning, and deep learning technologies that are all rooted in the types of data and the kinds of data structures we see in scientific disciplines. We specifically engage with scientific disciplines like ecology, environmental science, and earth observation, which address broad societal and environmental challenges we're facing today such as biodiversity loss, human-wildlife conflict, and ecosystem stability in the face of climate change.
I've been working at the intersection of computer vision and environmental conservation for about 13 years now. My early work focused on how to use computer vision to process ecological data. This work often included computer vision processes tailored for specific ecological tasks. For example, one project was developed to build up representative data structures for the spot patterns of individual snow leopards. From there, we would try to build reference systems to enable users to find mathematical matches to the relative geometries and shapes of these spots. I have subsequently watched the explosion of deep learning approaches for computer vision, and how technology shifts have changed how researchers in a number of applications thought about automating the extraction of scientific insight from image data, video data, or any sort of structured input data. This was important in two major ways. The first was the growing efficiency of computer vision technologies and thus decreasing costs required to run them. The second was a shift in computer vision work around data sharing and curating large datasets across many potential end users of a given system. Data sharing in its own right has introduced additional complexity to the application design process by raising questions around data sovreignty, data ownership, and when sharing data does or does not actually benefit the end user.
Something I'm really excited about right now is work which leverages and improves on the power of multimodal machine learning—a process that brings together a combination of data types into a deep learning model. These models can learn from not just one modality of data in isolation, but actually share information across modalities. In other words, the model can take advantage of existing data models in the aggregate to generate searchable data covering large natural world and environmental datasets. This is what can be considered “data bycatch”—the process of retrieving information from datasets (for example images or videos) that the collector of the data didn’t intend to capture but might be useful for another researcher or to answer a different and new ecological question. An example of this work can be shown in a collaboration with iNaturalist, a large community science platform for ecological data collection cataloging user images of things they have seen in the wild (plants, animals, insects) which get identified as species by the platform community. These images are then translated into research-grade observations in the Global Biodiversity Information Facility. But, as it turns out, every photo has more information captured in its pixels than just the species or site of interest, for example information about environmental conditions, species interactions, or animal behavior. For this reason, we are building systems that enable ecologists or interested community members to better query that data and pull out specific data of interest without needing to train a custom machine learning model for every possible goal.
Direct collaboration with iNaturalist, or the U.S. Fish and Wildlife Service and the community platform WildLabs, help to identify a diverse set of end users from researchers to amateur conservation groups. Their interests shape the kinds of projects we prioritize in the lab. I strongly believe it doesn't make sense to build a tool you think someone wants, and that it makes more sense to build a tool someone has asked for.
To me, community building is a necessity for successful interdisciplinary science. Regardless of the disciplines you're bringing together, if there exists a void between two disciplines where nobody is talking, and there are no mechanisms, spaces, or existing relationships to work together, there ends up being a huge barrier to progress within each respective field. In the communities I have helped to foster, I was driven by the fact that I didn't have access or insight, or was finding out about opportunities much later than I could take advantage of them. Because there weren't spaces to share information and to share best practices across disciplines despite the fact that there were all these individuals who were somewhat isolated at their own home institutions and were passionate about some of the same issues intersecting in the work that has driven my research to date: AI and conservation. So I started a Slack community called “AI for Conservation” a couple of years ago. Today, that slack community has grown to over 2000 researchers worldwide, with members from a really diverse set of backgrounds and with a really diverse set of interests.
I feel it's an odd framing to call public interest work an “obligation” where, for me, it's the motivation and longstanding goal behind our work in the Conservation AI community. I didn't become a computer vision researcher because I was necessarily inspired by computer vision as a set of interesting mathematical or technological tools—though I do really appreciate that—and I genuinely love it. More specifically, I was always motivated to do research that could of service to my community and to those who are in community with me, be it collaborators or students who join my group. I am deeply interested in how developments in the technology we build get translated into things that are accessible and can be used by people beyond our field of expertise. I translate this motivation to translate AI toward the public into my practice through partnerships and collaborations with organizations interested both in using the technology and have the technical capacity to try to do the translation, whether that's governmental agencies or some conservation organizations.
Building off these strong community-building experiences, and coming into MIT, one of the school’s limitations, I believe, is that it's a Technology Institute. At MIT, much like other Tech Institutes in the U.S., we don't have a broader base of different scientific disciplines, including departments focused on applied environmental disciplines such forestry, fisheries, or community ecology. We also lack representation from many of the social sciences.
I want to note that there are many people who do go out of their way to bring in, bridge, and build the collaborations with those types of scientists disciplines outside MIT, at local Boston level as well as across the country and the world. However, it is important to incentivize bridging gaps to scientific and social science disciplines that are not well represented at MIT. Bridging expertise is a strong mechanism to break down disciplinary silos not just within our institution, but across institutions. I think it's not easy to do, but if MIT could work to facilitate greater recognition of our own community limitations, this recognition would also allow us as a community to understand our own value in collaborations and partnerships to address these grand societal challenges.
I think another huge limitation of public interest technology is how it is perceived to misalign with capitalist motives such as profitability. For this reason, the translation of existing technologies for public use is an important challenge to figure out. At MIT, the traditional translation of technology has been via entrepreneurship and start-ups. But entrepreneurship doesn't necessarily make sense for nonprofit, social good, or environmental / climate applications. And yet, I think that the translation of technology for public interests is precisely something MIT is posed to potentially do well. This could, for example, employ data scientists or software engineers to focus on building out tools to translate technologies currently under development or in-use at MIT for the public interest and to create sustainable (both economic and environmental) translations of those technologies.
Lastly, I would add, Public Interest Technology initiatives are not happening in isolation. I actually think that Boston University, the University of Chicago, and many other schools that have been investing directly into this field and the work that coincides with it. And I think there's a lot of value in that.
Sara Beery is an assistant professor at MIT EECS' Faculty of AI and Decision Making and CSAIL, and was previously a Visiting Researcher at Google working on Auto Arborist. Sara has always loved the natural world, and she has seen a growing need for technology-based approaches to conservation and sustainability challenges. Beery’s research focuses on building computer vision methods that enable global-scale environmental and biodiversity monitoring across data modalities, tackling real-world challenges including strong spatiotemporal correlations that lead to domain shift, imperfect data quality, fine-grained categories, and long-tailed distributions. She received my PhD in Computing and Mathematical Sciences (CMS) at Caltech, advised by Pietro Perona, where she received the Amori Doctoral Prize for my dissertation. She was awarded both the PIMCO Data Science Fellowship and the Amazon AI4Science Fellowship, which recognize senior graduate students that have had a remarkable impact in machine learning and data science, and in their application to fields beyond computer science. Her work has been funded in part by an NSF Graduate Research Fellowship and the Caltech Resnick Sustainability Institute.
