Written by Audrey Woods
There is widespread agreement among scientists that we are in the midst of the sixth mass extinction event of biodiversity worldwide. Current extinction rates are faster than the last extinction of the dinosaurs with no signs of slowing down. CSAIL Assistant Professor Sara Beery says this issue is more urgent than many people recognize, arguing that human society “doesn’t understand how reliant it is on the stability of ecosystems and biodiversity.”
Considering the very short time horizons and the damage already done, there is a pressing need for policymakers, ecologists, and global communities to work toward stabilizing at-risk ecosystems. But biological systems are enormously complicated and it’s not always clear what actions would be most effective. To draw causal connections and understand the roots of the changes we’re seeing, it’s important to get robust and up-to-date data into the hands of those with the ability to act.
For that reason, Professor Beery is working on computer science methods that will help scientists approach the giant job of monitoring global biodiversity at scale, tackling real-world challenges such as adapting models for distribution shifts, object identification in visual data, and more. She is also passionate about giving ecologists, biologists, and non-computer scientists the tools to create their own AI solutions, with the aim of leveraging technology to improve and accelerate conservation efforts around the globe.
FINDING HER INTEREST
Professor Beery says she was initially motivated by the broader idea of using technology as a mechanism to address large scale social and environmental challenges, recognizing that “there is no way that we are going to be able to access the information needed about the environment to be able to make informed decisions without bringing in automated methods that enable us to scale.” Soon after starting her undergraduate degree at Seattle University, she got involved in a project using computer vision to identify individual snow leopards, which she says, “opened my eyes to computer vision as a specific tool that could be very impactful in these areas.” This led her to the California Institute of Technology, where she studied in the Computer Vision Lab and wrote her PhD thesis titled, “Where the Wild Things Are: Computer Vision for Global-Scale Biodiversity Monitoring.” Now, as a member of the MIT faculty, she continues to work on many of the challenges outlined there, improving methods to enable automated data processing and developing efficient human-AI systems for conservation.
Over the course of her academic career, Professor Beery has seen rapid expansion in this subfield of computer vision. “When I first started my PhD,” she jokes, “people thought I was crazy. I heard comments like, ‘no one cares about animals, it's too niche, why don’t you work on important applications like self-driving cars or robotics?’ The community was completely unfamiliar with both the importance of biodiversity and the fundamental technical challenges it presents for AI.” But her persistence paid off because, in part due to her community-building efforts, there are now thousands of scientists using and studying machine learning in ecology. For example, Professor Beery started a Slack channel called “AI for Conservation” that now has close to three thousand members around the world. “At first I knew everyone working in this area. Now the field has grown so much that I discover people who are working on problems that I helped shape and define who I've never met.”
The central challenge of ecological conservation that Professor Beery and her colleagues aim to address is enabling ecosystem monitoring to scale—geospatially, temporally, and taxonomically. On the one hand, there is an overload of raw data, with wildlife cameras, bioacoustic sensors, drones, GPS collars, and satellites gathering terabytes of data which historically had to be processed manually. Professor Beery says, “there are so few species that we have comprehensive enough information for to be able to even start scraping the surface of making these informed causal inferences on how this complex system is changing.” One way to bridge this gap and use the available data to gain insight into ecological niches and granular complexity is technology. That’s where Professor Beery’s research comes in.
NEW TOOLS AND METHODS FOR CONSERVATION
One particular hurdle when deploying ML models, especially in ecological applications, is what’s called distribution shift, or situations where the training data is different in some way from the data the model is deployed to predict on. As Professor Beery explains, “machine learning models really struggle to adapt to change,” which, unfortunately, is a necessary aspect of wildlife research as the world is constantly changing—and rates of climate and anthropogenic change are increasing. Changes in season, location, or population can degrade model performance and limit their real-world utility. To address this, Professor Beery and her colleagues have been applying unsupervised domain adaptation, “developing methods that learn from data from the new place you want the model to work, without requiring hand-labeled training data from that new place.” These methods can help researchers quickly adapt existing models to new environments—such as unseen test rivers while counting fish to estimate their escapement—without time-consuming and expensive retraining. This work, which offers an open-source package to help scientists apply domain adaptation in the field, is a continuation of Caltech Camera Traps and WILDS, both benchmarks that enable the AI community to measure progress on challenges posed by real-world distribution shifts.
Another project Professor Beery is excited about is recent work with iNaturalist, a social network platform for community sourced and labeled biodiversity images. With over 200 million photos associated with species observations in the wild, this database offers untold scientific potential for a variety of ecological questions. To help researchers and interested practitioners access this data, Professor Beery and her colleagues are incorporating improved vision language models to increase the ability for scientists to interact with and discover information captured in these images beyond their species label, expanding the platform’s scientific capacity. For instance, a user might be able to query the system for examples of a specific species of jellyfish washed up on beaches, or images of caterpillars with everted osmeterium (a technical term for when the defensive organ of certain caterpillar species are on display). Facilitating such specific scientific questions helps ecologists better leverage the iNaturalist platform for biological insight. “What we're working on now,” Professor Beery says, “is making the system even more interactive,” allowing scientists to give feedback and access the data they need even faster. Relatedly, Professor Beery’s group is expanding this research with the Smithsonian and the Snapshot USA project, designing models that enable scientists to efficiently interact with these centralized camera trap data repositories to “move beyond species and look at interactions, behavior, all these other dimensions of information that are captured in pixels but haven’t previously been accessible.”
While her research is focused on designing new tools and methods, Professor Beery is also passionate about educating non-experts in using machine learning to solve their own specific ecological problems. She says, “there’s a big bottleneck in terms of the capacity to apply the technologies effectively,” which has inspired her to create educational opportunities to bring people in from other disciplines and give them the skills to apply AI themselves. For example, she leads an educational program called CV4Ecology that, in 2025, will be going into its third year of training ecologists to apply computer vision in their practices. With these efforts, Professor Beery hopes to lower the bar to entry when it comes to creating specific AI and ML solutions to environmental challenges.
LOOKING FORWARD: ACADEMIA’S ROLE AND NEW IDEAS
Generally, Professor Beery finds the growth in this area encouraging, especially in academia. Pointing to the faculty searches that are happening at MIT and other institutions that explicitly combine ecology and computer science, she says, “it’s starting to be recognized as a unique subdiscipline, like AI for healthcare.” Compounding this growth, the rise of international regulations around biodiversity impact will create an ongoing need for better AI and ML tools to process, label, and study the ecological effects of business activities.
However, one thing Professor Beery would like to see more of is a conversation about how these tools can be translated from research institutions to the users who need them. Profit-driven entrepreneurialism, which has traditionally filled this role, is unlikely to drive the changes needed in this area and create the infrastructure to support biologists, ecologists, and other researchers in their work. Professor Beery imagines “hubs that are focused on translating public interest technology, maybe supported by foundations or governments.” This is a challenge for all climate technology, since, Professor Beery says, “it's not as immediate. It's hard to get people to look beyond yearly profit margins to what's potentially going to be incredibly impactful for not just your company, but all of society in the next fifty to one-hundred years.”
But, thanks to emerging policies and increasing public awareness, companies around the world will soon need accurate and efficient ways to evaluate their biodiversity impact. As consumers become more environmentally conscious and scientists across disciplines realize the benefits of applying computer science tools in their research, Professor Beery will continue to work on advancing technology solutions for conservation and sustainability.
To learn more about Professor Beery’s work, visit her website or CSAIL page.