Sara Beery

Elephant populations are under severe threat in the Greater Mara ecosystem, along the border between Kenya and Tanzania. The park wildlife managers who struggle to find the best measures to protect the giant mammals now can employ a “participatory AI” system that analyzes huge volumes of photos gathered by camera traps to keep track of individual animals—helping to assess the success of the protective strategies.

Participatory AI systems aim to “get the best of both worlds; to use AI for what it's good for, and bring in human expertise for what AI struggles at,” says Sara Beery, an MIT assistant professor of AI and decision making.

Beery's lab partnered with the Mara Elephant Project to develop a system that identifies each elephant based on the veins and contours of its ears, plus deep learning representations of its face. With collaborators at the University of Michigan, she and her colleagues can then strengthen the identifications by validating them with human experts.

“We can monitor an incredibly conflicted transnational elephant population of close to 2,000 individuals using this combination of AI and efficiently incorporating human expertise,” Beery says. “This was just was not possible before.”

Such tools address a critical worldwide need to understand rapid changes in ecosystems. “You need to be able to iterate quickly; you can't wait five years to find out that your management strategy was bad and the elephants are still in decline,” she says.

Beery works to bridge the gap between today's rapidly advancing computer technologies and dramatic global environmental challenges, with a particular focus on ecosystem stability and biodiversity loss. She partners with academic institutions, government agencies, non-governmental organizations and global research giants such as Google and IBM. She's also hearing from many corporate sustainability groups that recognize that they need to up their games to address the climate crisis.

Better guides to biodiversity
“The World Wide Fund for Nature estimates that we've lost almost 70% of all wild animals on earth since the 1970s,” she says. “So we have to do something, but it's not super clear what we need to do, and that's what I would like to figure out.”

With dramatic advances in cameras, satellite imaging, acoustic sensors, drone-based surveys, animal tracking devices and other ecosystem sensing equipment, scientists are collecting a huge explosion of wildlife data that can help to understand this global crisis.

However, “the vast majority of that data is sitting in hard drives under someone's desk untouched,” Beery says, “because we do not have the human expertise and capacity that we need to get the scientific insight out of these raw forms of data efficiently and cost-effectively.”

Her research has shown that AI can help in filtering through this enormous flood of ecological data to discover ecosystem trends and biodiversity losses. “We have to figure out how to make use of a combination of expert human intelligence and large-scale, hopefully very robust machine learning systems,” Beery says.

To complicate her task, she points out, environmental and ecological data are structured very differently than the data more commonly used in machine learning, such as images scraped from the Web.

Her studies also address a widely recognized general challenge in machine learning, which assumes that the data sets used to train a system are extremely similar to the data sets on which the systems will be applied. “That assumption is completely violated when you move to many real-world problems,” she notes.

For instance, consider AI models trained on the 11 million photos taken by her network of camera traps in East African parks each year. With climate change, human encroachment and other factors at play, “models that are trained this year don't work next year,” Beery says. The models need to be made robust enough to adapt.

Her lab also builds AI systems to support a broad spectrum of ecological science. For example, a partnership with the University of Edinburgh and the University of Massachusetts at Amherst taps into iNaturalist, a community science platform that includes nearly 200 million species observations from more than 450,000 species.

The iNaturalist project aims to develop large language model text queries that incorporate scientific expertise, so that an ecologist can easily explore a hypothesis with existing data. For instance, the scientist might want to know exactly which flowers one type of hummingbird is eating across the Pacific Northwest. The system potentially could examine all the images where the hummingbird is feeding, identify the flowers and summarize these findings.

We're moving from this goal of being 'carbon net zero' to being 'nature net zero'

Industry looks towards 'nature net zero'
Beery is collaborating with IBM and Google DeepMind to examine how their large-scale geospatial models can assist in monitoring biodiversity, particularly from space. She also studies how such large-scale models can be integrated with ground-level data gathered by conservation organizations.

Additionally, she's seeing a clear and expanding need for corporations to develop biodiversity impact assessments. “We're moving from this goal of being 'carbon net zero' to being 'nature net zero',” Beery says.

Companies and agencies such as the World Bank are actively investing in biodiversity assessments, to examine strategies to offset ecosystem damage and to make the best design choices for extracting materials or manufacturing products. This trend is being driven both by tough new international regulations and by the need for corporations to show their shareholders that they are working to reduce environmental damage.

Interdisciplinary work on big problems
Beginning at age 16, Beery spent six years as a professional ballet dancer. She always kept in mind the need to forge a new career when her dance adventure came to a close. “I knew that when I retired from ballet, I wanted to try to devote myself to the large societal problems facing us around climate change,” she says.

She tells a story about being initially drawn to scientific talks at the Georgia Institute of Technology by the free food, because she was being paid very little by the Atlanta Ballet. “Those talks introduced me to the idea that engineering and technology could be a force for social change, social justice and environmental justice,” she says.

She decided to pursue training in science and technology that could allow her to make significant contributions. “Framing technology as a force for societal benefit really helped me fall in love with the math itself,” she adds.

Beery earned a PhD in computing and mathematical sciences at the California Institute of Technology. She joined Google Research for a year, studying advanced methods to map urban trees, before coming to MIT.

She was particularly drawn to the Institute by the Schwarzman College of Computing's mission of addressing grand challenges for the planet with a cadre of experts who are deeply serious about interdisciplinary cooperation.

True interdisciplinary engagement means “directly engaging and working together with different scientific disciplines, spending the time to gain expertise in other scientific disciplines, and then figuring out how do we actually solve these problems, and not just write papers about them but build systems that can be robustly, efficiently and actively deployed,” she emphasizes. “MIT is a place making that a core priority.”

One of her long-term goals is to work toward early-warning earth monitoring systems that can flag ecosystem concerns in real time and high detail. That way, “we're not just reacting after species have gone extinct, or after we're seeing things like the massive scale collapse of the snow crab fishery last year in Alaska,” Beery says.

“How do we recognize that this will happen, and then what can we do?” she asks. “What are the mitigating actions that we can take ahead of time, to be a little bit more proactive instead of reactive in our management strategies for the Earth?”