For more than 30 years, the MacVicar Faculty Fellows Program has recognized exemplary and sustained contributions to undergraduate education at MIT. The program is named in honor of Margaret MacVicar, MIT’s first dean for undergraduate education and founder of the Undergraduate Research Opportunities Program (UROP). Fellows are chosen through an annual and highly competitive nomination process. The Registrar’s Office coordinates and administers the award on behalf of the Division of Graduate and Undergraduate Education. Nominations are reviewed by an advisory committee, and the provost selects the fellows.

Amos Winter: Bringing excitement to the classroom

Amos Winter is the Germeshausen Professor in the Department of Mechanical Engineering (MechE). He joined the faculty in 2012 and is best known for teaching class 2.007 (Design and Manufacturing I).

A hallmark of Winter’s pedagogy is the way he connects technical learning and core engineering science with real-world impacts. His approach keeps students actively engaged and encourages critical thinking while developing their competence and confidence as design engineers. Current graduate student Ariel Mobius ’24 writes, “Professor Winter is a transformative educator. He successfully blends rigorous technical instruction with lessons on problem scoping and hands-on learning and backs it all up with personalized mentorship. He is a committed advocate for his students and has fundamentally shaped my path as a mechanical engineer.”

Especially notable is Winter’s energetic style and use of interactive materials and demonstrations to make fundamental topics tangible. “He wheels in a large steamer trunk filled with demos he has built or collected to illustrate the day’s topic,” writes Class of 1948 Career Development Professor and assistant professor of mechanical engineering Kaitlyn Becker. “Some demos are enduring classics and others newly designed each year.” Through his “Gearhead Moment of Zen” Winter will share an astonishing car stunt to explain the mechanics using course material. “The theatrics stay in students’ minds,” says Becker, highlighting how Winter’s dramatic examples reinforce learning.

These techniques, combined with a supportive culture, allowed Winter to transform 2.007 from a core class and first subject in engineering design into a celebration of student effort and learning. Throughout the term, students learn how to design and build objects culminating in a robot competition in which their creations tackle themed challenges on a life-size game board. In the past, fewer than half the students were able to compete and today, boosted by Winter’s mentorship and enthusiasm, nearly 97 percent finish a competition-ready robot.

Ralph E. and Eloise F. Cross Professor of Mechanical Engineering David Hardt writes, “Thanks to Amos, this subject has become transformative for many MechE undergraduates.” Becker concurs: “He is the heart and captain of the 2.007 ‘cheer squad,’ cultivating a caring and motivated teaching team.”

Current graduate student Aidan Salazar ’25 notes, “His teaching philosophy is grounded in empowerment: he encourages students to take risks when designing while giving them the confidence and support needed to do so with thoughtful engineering analysis.”

Winter is also deeply invested in students’ growth outside the classroom. He serves as faculty supervisor for MIT’s Formula SAE (Society of Automotive Engineers) and Solar Car teams and guides related UROP projects. In fall 2025 alone, he advised nearly 50 UROP students from the teams, demonstrating his commitment to experiential learning and ability to mentor students at scale.

Salazar continues: “He has offered extraordinary contributions in helping MIT undergraduates embody the Institute’s ‘mens-et-manus’ [‘mind-and-hand’] motto, and I am grateful to be one of the individuals shaped by his teaching.”

“I have always looked up to my colleagues who are MacVicar Fellows as the best educators at the Institute,” writes Winter. “What makes this acknowledgement even more special to me is by earning it from teaching 2.007, which I often cite as one of the best parts of my job. The class is where most mechanical engineering undergraduates gain their first real engineering experience by physically realizing a machine of their own conception. It has been extremely gratifying to watch a generation of students translate their knowledge of engineering and design from the class into their careers … I am honored to have played a role in their intellectual growth and done so meaningfully enough to be recognized as a MacVicar Fellow.”

Nickolai Zeldovich: Inspiring independent thinkers and future teachers

Nickolai Zeldovich is the Joan and Irwin M. (1957) Jacobs Professor of Electrical Engineering and Computer Science (EECS). Student testimonials highlight his unique ability to activate their problem-solving skills, cultivate their intellectual curiosity, and infuse learning with joy.

Katarina Cheng ’25 writes, “From my first day of lecture in the course, I was immediately drawn in by Professor Zeldovich’s joy and enthusiasm for every facet of security and its power,” and Rotem Hemo ’17, ’18 says that Zeldovich “empowers students to find solutions themselves.”

Yael Tauman Kalai, the Ellen Swallow Richards (1873) Professor and professor of EECS concurs. She notes that his lectures — with back-and-forth discussion and probing questions — encourage independent thinking and ensure that “everyone feels a little smarter at the end. It is not surprising that students love him.”

Zeldovich’s affinity for problem-solving translates to his curricular work as well. When he arrived at MIT in 2008, Course 6 offered classes in theoretical and applied cryptography, but lacked a dedicated systems security subject. Recognizing this as a significant gap, Zeldovich took it upon himself to create class 6.566/6.858 (Computer Systems Security) in 2009. Since then, the subject has become a central part of the curriculum, but sustained interest from undergraduates revealed another need, and in 2021 he partnered with colleagues to create a dedicated introductory course: 6.1600 (Foundations of Computer Security).

Edwin Sibley Webster Professor of EECS Srini Devadas writes: “What our curriculum was sorely in need of was a systems security class, and Nickolai immediately and single-handedly created [it],” and has “taught this class to rave reviews ever since.”

The impact of Zeldovich’s thoughtful, inquiry-driven approach to pedagogy extends beyond the walls of his classroom, inspiring future educators, teaching assistants (TAs), and even his faculty colleagues at MIT.

Henry Corrigan-Gibbs, the Douglas Ross (1954) Career Development Professor of Software Technology and associate professor of computer science, writes that Zeldovich has “proven himself to be a dedicated teacher of teachers … One of the things that makes teaching with Nickolai so much fun is that he shares his passion with the undergraduates and MEng students who join the course staff as TAs.”

“[He] encourages the TAs to contribute their own creative ideas to the course,” continues Corrigan-Gibbs. “It should not be a surprise then that 100% of the TAs that we have had in our class have signed up to teach with Nickolai again.”

“Due, in no small part, to how I saw Nickolai lead his classroom, I was inspired to become an educator myself,” writes MIT alumna Anna Arpaci-Dusseau ’23, SM ’24. “I saw that the role of an instructor is not only to teach, but to innovate by thinking of creative projects, and to connect by listening to students’ concerns. As I go forward in my career, I am grateful to have such a wonderful example of an educator to look up to.”

Kalai adds, “I have learned a great deal from the two times that I have ‘taken’ (part of) the class from Nickolai. His extensive knowledge and experience are evident in every lecture. There is so much variety to Nickolai’s teaching.”

Nickolai Zeldovich is the recipient of numerous awards including the EECS Spira Teaching Award (2013), the Edgerton Faculty Achievement Award (2014), the EECS Faculty Research Innovation Fellowship (2018), and the EECS Jamieson Award for Excellence in Teaching (2024).

On receiving this award, Zeldovich says, “MIT has a culture of strong undergraduate education, so being selected as a MacVicar Fellow was truly an honor. It’s a joy to teach smart students about computer systems, and the tradition of co-teaching classes in the EECS department helped me improve as a teacher. Most of all, I look forward to continuing to teach MIT’s students!”

Learn more about the MacVicar Faculty Fellows Program on the Registrar’s Office website.

The post 2026 MacVicar Faculty Fellows named appeared first on MIT Schwarzman College of Computing.

In the 21st century, the management of heart failure has become decidedly less medieval: Today, patients undergo a combination of healthy lifestyle changes, prescription of medications, and sometimes use pacemakers. Yet heart failure remains one of the leading causes of morbidity and mortality, placing a substantial burden on health-care systems across the globe. 

“About half of the people diagnosed with heart failure will die within five years of diagnosis,” says Teya Bergamaschi, an MIT PhD student in the lab of Nina T. and Robert H. Rubin Professor Collin Stultz and the co-first author of a new paper introducing a deep learning model for predicting heart failure. “Understanding how a patient will fare after hospitalization is really important in allocating finite resources.”

The paper, published in Lancet eClinical Medicine by a team of researchers at MIT, Mass General Brigham, and Harvard Medical School, shares results from developing and testing PULSE-HF, which stands loosely for “Predict changes in left ventricULar Systolic function from ECGs of patients who have Heart Failure.” The project was conducted in Stultz’s lab, which is affiliated with the MIT Abdul Latif Jameel Clinic for Machine Learning in Health. Developed and retrospectively tested across three different patient cohorts from Massachusetts General Hospital, Brigham and Women’s Hospital, and MIMIC-IV (a publicly available dataset), the deep learning model accurately predicts changes in the left ventricular ejection fraction (LVEF), which is the percentage of blood being pumped out of the left ventricle of the heart.

A healthy human heart pumps out about 50 to 70 percent of blood from the left ventricle with each beat — anything less is considered a sign of a potential problem. “The model takes an [electrocardiogram] and outputs a prediction of whether or not there will be an ejection fraction within the next year that falls below 40 percent,” says Tiffany Yau, an MIT PhD student in Stultz’s lab who is also co-first author of the PULSE-HF paper. “That is the most severe subgroup of heart failure.” 

If PULSE-HF predicts that a patient’s ejection fraction is likely to worsen within a year, the clinician can prioritize the patient for follow-up. Subsequently, lower-risk patients can reduce their number of hospital visits and the amount of time spent getting 10 electrodes adhered to their body for a 12-lead ECG. The model can also be deployed in low-resource clinical settings, including doctors offices in rural areas that don’t typically have a cardiac sonographer employed to run ultrasounds on a daily basis.

“The biggest thing that distinguishes [PULSE-HF] from other heart failure ECG methods is instead of detection, it does forecasting,” says Yau. The paper notes that to date, no other methods exist for predicting future LVEF decline among patients with heart failure.

During the testing and validation process, the researchers used a metric known as “area under the receiver operating characteristic curve” (AUROC) to measure PULSE-HF’s performance. AUROC is typically used to measure a model’s ability to discriminate between classes on a scale from 0 to 1, with 0.5 being random and 1 being perfect. PULSE-HF achieved AUROCs ranging from 0.87 to 0.91 across all three patient cohorts.

Notably, the researchers also built a version of PULSE-HF for single-lead ECGs, meaning only one electrode needs to be placed on the body. While 12-lead ECGs are generally considered superior for being more comprehensive and accurate, the performance of the single-lead version of PULSE-HF was just as strong as the 12-lead version.

Despite the elegant simplicity behind the idea of PULSE-HF, like most clinical AI research, it belies a laborious execution. “It’s taken years [to complete this project],” Bergamaschi recalls. “It’s gone through many iterations.” 

One of the team’s biggest challenges was collecting, processing, and cleaning the ECG and echocardiogram datasets. While the model aims to forecast a patient’s ejection fraction, the labels for the training data weren’t always readily available. Much like a student learning from a textbook with an answer key, labeling is critical for helping machine-learning models correctly identify patterns in data.

Clean, linear text in the form of TXT files typically works best when training models. But echocardiogram files typically come in the form of PDFs, and when PDFs are converted to TXT files, the text (which gets broken up by line breaks and formatting) becomes difficult for the model to read. The unpredictable nature of real-life scenarios, like a restless patient or a loose lead, also marred the data. “There are a lot of signal artifacts that need to be cleaned,” Bergamaschi says. “It’s kind of a never-ending rabbit hole.”

While Bergamaschi and Yau acknowledge that more complicated methods could help filter the data for better signals, there is a limit to the usefulness of these approaches. “At what point do you stop?” Yau asks. “You have to think about the use case — is it easiest to have this model that works on data that is slightly messy? Because it probably will be.”

The researchers anticipate that the next step for PULSE-HF will be testing the model in a prospective study on real patients, whose future ejection fraction is unknown.

Despite the challenges inherent to bringing clinical AI tools like PULSE-HF over the finish line, including the possible risk of prolonging a PhD by another year, the students feel that the years of hard work were worthwhile. 

“I think things are rewarding partially because they’re challenging,” Bergamaschi says. “A friend said to me, ‘If you think you will find your calling after graduation, if your calling is truly calling, it will be there in the one additional year it takes you to graduate.’ … The way we’re measured as researchers in [the ML and health] space is different from other researchers in ML space. Everyone in this community understands the unique challenges that exist here.”

“There’s too much suffering in the world,” says Yau, who joined Stultz’s lab after a health event made her realize the importance of machine learning in health care. “Anything that tries to ease suffering is something that I would consider a valuable use of my time.” 

The post Can AI help predict which heart-failure patients will worsen within a year? appeared first on MIT Schwarzman College of Computing.

Their method uses a specialized vision-language model to perceive the scenario in an image and simulate actions needed to reach a goal. Then a second model translates those simulations into a standard programming language for planning problems, and refines the solution.

In the end, the system automatically generates a set of files that can be fed into classical planning software, which computes a plan to achieve the goal. This two-step system generated plans with an average success rate of about 70 percent, outperforming the best baseline methods that could only reach about 30 percent.

Importantly, the system can solve new problems it hasn’t encountered before, making it well-suited for real environments where conditions can change at a moment’s notice.

“Our framework combines the advantages of vision-language models, like their ability to understand images, with the strong planning capabilities of a formal solver,” says Yilun Hao, an aeronautics and astronautics (AeroAstro) graduate student at MIT and lead author of an open-access paper on this technique. “It can take a single image and move it through simulation and then to a reliable, long-horizon plan that could be useful in many real-life applications.”

She is joined on the paper by Yongchao Chen, a graduate student in the MIT Laboratory for Information and Decision Systems (LIDS); Chuchu Fan, an associate professor in AeroAstro and a principal investigator in LIDS; and Yang Zhang, a research scientist at the MIT-IBM Watson AI Lab. The paper will be presented at the International Conference on Learning Representations.

Tackling visual tasks

For the past few years, Fan and her colleagues have studied the use of generative AI models to perform complex reasoning and planning, often employing large language models (LLMs) to process text inputs.

Many real-world planning problems, like robotic assembly and autonomous driving, have visual inputs that an LLM can’t handle well on its own. The researchers sought to expand into the visual domain by utilizing vision-language models (VLMs), powerful AI systems that can process images and text.

But VLMs struggle to understand spatial relationships between objects in a scene and often fail to reason correctly over many steps. This makes it difficult to use VLMs for long-range planning.

On the other hand, scientists have developed robust, formal planners that can generate effective long-horizon plans for complex situations. However, these software systems can’t process visual inputs and require expert knowledge to encode a problem into language the solver can understand.

Fan and her team built an automatic planning system that takes the best of both methods. The system, called VLM-guided formal planning (VLMFP), utilizes two specialized VLMs that work together to turn visual planning problems into ready-to-use files for formal planning software.

The researchers first carefully trained a small model they call SimVLM to specialize in describing the scenario in an image using natural language and simulating a sequence of actions in that scenario. Then a much larger model, which they call GenVLM, uses the description from SimVLM to generate a set of initial files in a formal planning language known as the Planning Domain Definition Language (PDDL).

The files are ready to be fed into a classical PDDL solver, which computes a step-by-step plan to solve the task. GenVLM compares the results of the solver with those of the simulator and iteratively refines the PDDL files.

“The generator and simulator work together to be able to reach the exact same result, which is an action simulation that achieves the goal,” Hao says.

Because GenVLM is a large generative AI model, it has seen many examples of PDDL during training and learned how this formal language can solve a wide range of problems. This existing knowledge enables the model to generate accurate PDDL files.

A flexible approach

VLMFP generates two separate PDDL files. The first is a domain file that defines the environment, valid actions, and domain rules. It also produces a problem file that defines the initial states and the goal of a particular problem at hand.

“One advantage of PDDL is the domain file is the same for all instances in that environment. This makes our framework good at generalizing to unseen instances under the same domain,” Hao explains.

To enable the system to generalize effectively, the researchers needed to carefully design just enough training data for SimVLM so the model learned to understand the problem and goal without memorizing patterns in the scenario. When tested, SimVLM successfully described the scenario, simulated actions, and detected if the goal was reached in about 85 percent of experiments.

Overall, the VLMFP framework achieved a success rate of about 60 percent on six 2D planning tasks and greater than 80 percent on two 3D tasks, including multirobot collaboration and robotic assembly. It also generated valid plans for more than 50 percent of scenarios it hadn’t seen before, far outpacing the baseline methods.

“Our framework can generalize when the rules change in different situations. This gives our system the flexibility to solve many types of visual-based planning problems,” Fan adds.

In the future, the researchers want to enable VLMFP to handle more complex scenarios and explore methods to identify and mitigate hallucinations by the VLMs.

“In the long term, generative AI models could act as agents and make use of the right tools to solve much more complicated problems. But what does it mean to have the right tools, and how do we incorporate those tools? There is still a long way to go, but by bringing visual-based planning into the picture, this work is an important piece of the puzzle,” Fan says.

This work was funded, in part, by the MIT-IBM Watson AI Lab.

The post A better method for planning complex visual tasks appeared first on MIT Schwarzman College of Computing.

Today, artificial intelligence and science find themselves at a similar inflection point. The current AI revolution has been fueled by decades of research in the mathematical and physical sciences (MPS), which provided the challenging problems, datasets, and insights that made modern AI possible. The 2024 Nobel Prizes in physics and chemistry, recognizing foundational AI methods rooted in physics and AI applications for protein design, made this connection impossible to miss.

In 2025, MIT hosted a Workshop on the Future of AI+MPS, funded by the National Science Foundation with support from the MIT School of Science and the MIT departments of Physics, Chemistry, and Mathematics. The workshop brought together leading AI and science researchers to chart how the MPS domains can best capitalize on — and contribute to — the future of AI. Now a white paper, with recommendations for funding agencies, institutions, and researchers, has been published in Machine Learning: Science and Technology. In this interview, Jesse Thaler, MIT professor of physics and chair of the workshop, describes key themes and how MIT is positioning itself to lead in AI and science.

Q: What are the report’s key themes regarding last year’s gathering of leaders across the mathematical and physical sciences?

A: Gathering so many researchers at the forefront of AI and science in one room was illuminating. Though the workshop participants came from five distinct scientific communities — astronomy, chemistry, materials science, mathematics, and physics — we found many similarities in how we are each engaging with AI. A real consensus emerged from our animated discussions: Coordinated investment in computing and data infrastructures, cross-disciplinary research techniques, and rigorous training can meaningfully advance both AI and science.

One of the central insights was that this has to be a two-way street. It’s not just about using AI to do better science; science can also make AI better. Scientists excel at distilling insights from complex systems, including neural networks, by uncovering underlying principles and emergent behaviors. We call this the “science of AI,” and it comes in three flavors: science driving AI, where scientific reasoning informs foundational AI approaches; science inspiring AI, where scientific challenges push the development of new algorithms; and science explaining AI, where scientific tools help illuminate how machine intelligence actually works.

In my own field of particle physics, for instance, researchers are developing real-time AI algorithms to handle the data deluge from collider experiments. This work has direct implications for discovering new physics, but the algorithms themselves turn out to be valuable well beyond our field. The workshop made clear that the science of AI should be a community priority — it has the potential to transform how we understand, develop, and control AI systems.

Of course, bridging science and AI requires people who can work across both worlds. Attendees consistently emphasized the need for “centaur scientists” — researchers with genuine interdisciplinary expertise. Supporting these polymaths at every career stage, from integrated undergraduate courses to interdisciplinary PhD programs to joint faculty hires, emerged as essential.

Q: How do MIT’s AI and science efforts align with the workshop recommendations?

A: The workshop framed its recommendations around three pillars: research, talent, and community. As director of the NSF Institute for Artificial Intelligence and Fundamental Interactions (IAIFI) — a collaborative AI and physics effort among MIT and Harvard, Northeastern, and Tufts universities — I’ve seen firsthand how effective this framework can be. Scaling this up to MIT, we can see where progress is being made and where opportunities lie.

On the research front, MIT is already enabling AI-and-science work in both directions. Even a quick scroll through MIT News shows how individual researchers across the School of Science are pursuing AI-driven projects, building a pipeline of knowledge and surfacing new opportunities. At the same time, collaborative efforts like IAIFI and the Accelerated AI Algorithms for Data-Driven Discovery (A3D3) Institute concentrate interdisciplinary energy for greater impact. The MIT Generative AI Impact Consortium is also supporting application-driven AI work at the university scale.

To foster early-career AI-and-science talent, several initiatives are training the next generation of centaur scientists. The MIT Schwarzman College of Computing’s Common Ground for Computing Education program helps students become “bilingual” in computing and their home discipline. Interdisciplinary PhD pathways are also gaining traction; IAIFI worked with the MIT Institute for Data, Systems, and Society to create one in physics, statistics, and data science, and about 10 percent of physics PhD students now opt for it — a number that’s likely to grow. Dedicated postdoctoral roles like the IAIFI Fellowship and Tayebati Fellowship give early-career researchers the freedom to pursue interdisciplinary work. Funding centaur scientists and giving them space to build connections across domains, universities, and career stages has been transformative.

Finally, community-building ties it all together. From focused workshops to large symposia, organizing interdisciplinary events signals that AI and science isn’t siloed work — it’s an emerging field. MIT has the talent and resources to make a significant impact, and hosting these gatherings at multiple scales helps establish that leadership.

Q: What lessons can MIT draw about further advancing its AI-and-science efforts?

A: The workshop crystallized something important: The institutions that lead in AI and science will be the ones that think systematically, not piecemeal. Resources are finite, so priorities matter. Workshop attendees were clear about what becomes possible when an institution coordinates hires, research, and training around a cohesive strategy.

MIT is well positioned to build on what’s already underway with more structural initiatives — joint faculty lines across computing and scientific domains, expanded interdisciplinary degree pathways, and deliberate “science of AI” funding. We’re already seeing moves in this direction; this year, the MIT Schwarzman College of Computing and the Department of Physics are conducting their first-ever joint faculty search, which is exciting to see.

The virtuous cycle of AI-and-science has the potential to be truly transformative — offering deeper insight into AI, accelerating scientific discovery, and producing robust tools for both. By developing an intentional strategy, MIT will be well positioned to lead in, and benefit from, the coming waves of AI.

The post 3 Questions: On the future of AI and the mathematical and physical sciences appeared first on MIT Schwarzman College of Computing.

At MIT, a friendship between two professors — one an anthropologist, the other a computer scientist — led to creation of an undergraduate class that set out to find the answer to those questions. Combining the two seemingly disparate disciplines, the class encourages students to design artificial intelligence chatbots in humane ways that help users improve themselves.

The class, 6.S061/21A.S02 (Humane User Experience Design, a.k.a. Humane UXD), is an upper-level computer science class cross-listed with anthropology. This unique cross-listing allows computer science majors to fulfill a humanities requirement while also pursuing their career objectives. The two professors use methods from linguistic anthropology to teach students how to integrate the interactional and interpersonal needs of humans into programming.

Professor Arvind Satyanarayan, a computer scientist whose research develops tools for interactive data visualization and user interfaces, and Professor Graham Jones, an anthropologist whose research focuses on communication, created Humane UXD last summer with a grant from the MIT Morningside Academy for Design (MAD). The MIT MAD Design Curriculum Program provides funding for faculty to develop new classes or enhance existing classes using innovative pedagogical approaches that transcend departmental boundaries. Alongside the grant provided by MAD, Jones and Satyanarayan received funding to develop Humane UXD under the auspices of the Common Ground for Computing Education, an initiative of the MIT Schwarzman College of Computing that brings together departments to create courses integrating computing with other disciplines.

The Design Curriculum Program is currently accepting applications for the 2026-27 academic year; the deadline is Friday, March 20.

Jones and Satyanarayan met several years ago when they co-advised a doctoral student’s research on data visualization for visually impaired people. They’ve since become close friends who can pretty much finish one another’s sentences.

“There’s a way in which you don’t really fully externalize what you know or how you think until you’re teaching,” Jones says. “So, it’s been really fun for me to see Arvind unfurl his expertise as a teacher in a way that lets me see how the pieces fit together — and discover underlying commonalities between our disciplines and our ways of thinking.”

Satyanarayan continues that thought: “One of the things I really enjoyed is the reciprocal version of what Graham said, which is that my field — human-computer interaction — inherited a lot of methods from anthropology, such as interviews and user studies and observation studies. And over the decades, those methods have gotten more and more watered down. As a result, a lot of things have been lost.

“For instance, it was very exciting for me to see how an anthropologist teaches students to interview people. It’s completely different than how I would do it. With my way, we lose the rapport and connection you need to build with your interview participant. Instead, we just extract data from them.”

For Jones’ part, teaching with a computer scientist holds another kind of allure: design. He says that human speech and interaction are organized into underlying genres with stable sets of rules that differentiate an interview at a cocktail party from a conversation at a funeral.

“ChatGPT and other large language models are trained on naturally occurring human communication, so they have all those genres inside them in a latent state, waiting to be activated,” he says.

“As a social scientist, I teach methods for analyzing human conversation, and give students very powerful tools to do that. But it ends up usually being an exercise in pure research, whereas this is a design class, where students are building real-world systems.”

The curriculum appears to be on target for preparing students for jobs after graduation. One student sought permission to miss class for a week because he had a trial internship at a chatbot startup; when he returned, he said his work at the startup was just like what he was learning in class. He got the job.

The sampling of group projects below, built with Google’s Gemini, demonstrates some of what’s possible when, as Jones says, “there’s a really deep intertwining of the technology piece with the humanities piece.” The students’ design work shows that entirely new ways of programming can be conceptualized when the humane is made a priority.

The bots demonstrate clearly that an interdisciplinary class can be designed in such a way that everyone benefits: Students learn more and differently; they can fulfill a non-major course requirement by taking a class that is directly beneficial to their careers; and long-term faculty partnerships can be forged or strengthened.

Team Pond

One project promises to be particularly useful for graduating seniors. Pond is designed to help young college graduates adapt to the challenges of independent adult life. Team Pond configured the chatbot not to simply parrot the user, or to sycophantically praise wrong answers. Instead, Pond provides advice to help with “adulting” (behaving as a responsible adult).

“Pond is built to be your companion from college life into post-college life, to help you in your transition from being a small fish in a small pond to being a small fish in a very big pond,” says sophomore Mary Feliz.

“College is very much a high-proximity and high-context environment, in the sense that everybody around you is going through the same thing, and it’s easy to build relationships or find opportunities, because there are structured pathways that you have access to,” explains graduate student Emaan Khan. “Post-grad life is low-context. You’re not always surrounded by your peers or your professors. It’s no-proximity also, in the sense that you don’t have opportunities at your doorstep. Pond is a tool to help empower you to access certain opportunities, or learn how to navigate.”

Pond’s developers are graduate student Claire Camacho, Harvard University graduate student Charlotte Cheah, Feliz, and Khan. They trained Pond to offer expertise in three areas: social life, professional life, and adult skills.

The students demonstrated Pond’s social communication by asking how to start conversations in a hobby group. Next, they exhibited Pond’s adulting communication by typing in a question.

“My landlord has asked me personal questions that make me uncomfortable. What should I do?”

Appearing in Pond’s textbox was this response: “Politely set boundaries. You are not obligated to answer personal questions. You can gently redirect the conversation back to property-related topics. Know your rights: familiarize yourself with tenant rights in your city or state regarding privacy and landlord conduct.”

When asked how to do that, Pond offered specific instructions: “You can try saying phrases like: ‘I prefer to keep our conversations focused on the property,’ and ‘let’s stick to topics related to the apartment, please.’ Always redirect the conversation back to the apartment or a legitimate maintenance issue. Keep your tone polite but firm. Document any conversations if needed.”

Pond also offered a role-playing scenario to help the user learn what polite-but-firm language might be in that situation.

“The ethos of the practice mode is that you are actively building a skill, so that after using Pond for some time, you feel confident that you can swim on your own,” Khan says. The chatbot uses a point system that allows users to graduate from a topic, and a treasure chest to store prizes, elements added to boost the bot’s appeal.

Team News Nest

Another of the projects, News Nest, provides a sophisticated means of helping young people engage with credible news sources in a way that makes it fun. The name is derived from the program’s 10 appealing and colorful birds, each of which focuses on a particular area of news. If you want the headlines, you ask Polly the Parrot, the main news carrier; if you’re interested in science, Gaia the Goose guides you. The flock also includes Flynn the Falcon, sports reporter; Credo the Crow, for crime and legal news; Edwin the Eagle, a business and economics news guide; Pizzazz the Peacock for pop and entertainment stories; and Pixel the Pigeon, a technology news specialist.

News Nest’s development team is made up of MIT seniors Tiana Jiang and Krystal Montgomery, and junior Natalie Tan. They intentionally built News Nest to prevent “doomscrolling,” provide media transparency (sources and political leanings are always shown), and they created a clever, healthy buffer from emotional manipulation and engagement traps by employing birds rather than human characters.

Team M^3 (Multi-Agent Murder Mystery)

A third team, M^3, decided to experiment with making AI humane by keeping it fun. MIT senior Rodis Aguilar, junior David De La Torre, and second-year Deeraj Pothapragada developed M^3, a social deduction multi-agent murder mystery that incorporates four chatbots as different personalities: Gemini, OpenAI’s ChatGPT, xAI’s Grok, and Anthropic’s Claude. The user is the fifth player.

Like a regular murder mystery, there are locations, weapons, and lies. The user has to guess who committed the murder. It’s very similar to a board or online game played with real players, only these are enhanced AI opponents you can’t see, who may or may not tell the truth in response to questions. Users can’t get too involved with one chatbot, because they’re playing all four. Also, as in a real life murder mystery game, the user is sometimes guilty.

The post New MIT class uses anthropology to improve chatbots appeared first on MIT Schwarzman College of Computing.

Concept bottleneck modeling is one method that enables artificial intelligence systems to explain their decision-making process. These methods force a deep-learning model to use a set of concepts, which can be understood by humans, to make a prediction. In new research, MIT computer scientists developed a method that coaxes the model to achieve better accuracy and clearer, more concise explanations.

The concepts the model uses are usually defined in advance by human experts. For instance, a clinician could suggest the use of concepts like “clustered brown dots” and “variegated pigmentation” to predict that a medical image shows melanoma.

But previously defined concepts could be irrelevant or lack sufficient detail for a specific task, reducing the model’s accuracy. The new method extracts concepts the model has already learned while it was trained to perform that particular task, and forces the model to use those, producing better explanations than standard concept bottleneck models.

The approach utilizes a pair of specialized machine-learning models that automatically extract knowledge from a target model and translate it into plain-language concepts. In the end, their technique can convert any pretrained computer vision model into one that can use concepts to explain its reasoning.

“In a sense, we want to be able to read the minds of these computer vision models. A concept bottleneck model is one way for users to tell what the model is thinking and why it made a certain prediction. Because our method uses better concepts, it can lead to higher accuracy and ultimately improve the accountability of black-box AI models,” says lead author Antonio De Santis, a graduate student at Polytechnic University of Milan who completed this research while a visiting graduate student in the Computer Science and Artificial Intelligence Laboratory (CSAIL) at MIT.

He is joined on a paper about the work by Schrasing Tong SM ’20, PhD ’26; Marco Brambilla, professor of computer science and engineering at Polytechnic University of Milan; and senior author Lalana Kagal, a principal research scientist in CSAIL. The research will be presented at the International Conference on Learning Representations.

Building a better bottleneck

Concept bottleneck models (CBMs) are a popular approach for improving AI explainability. These techniques add an intermediate step by forcing a computer vision model to predict the concepts present in an image, then use those concepts to make a final prediction.

This intermediate step, or “bottleneck,” helps users understand the model’s reasoning.

For example, a model that identifies bird species could select concepts like “yellow legs” and “blue wings” before predicting a barn swallow.

But because these concepts are often generated in advance by humans or large language models (LLMs), they might not fit the specific task. In addition, even if given a set of pre-defined concepts, the model sometimes utilizes undesirable learned information anyway, which is a problem known as information leakage.

“These models are trained to maximize performance, so the model might secretly use concepts we are unaware of,” De Santis explains.

The MIT researchers had a different idea: Since the model has been trained on a vast amount of data, it may have learned the concepts needed to generate accurate predictions for the particular task at hand. They sought to build a CBM by extracting this existing knowledge and converting it into text a human can understand.

In the first step of their method, a specialized deep-learning model called a sparse autoencoder selectively takes the most relevant features the model learned and reconstructs them into a handful of concepts. Then, a multimodal LLM describes each concept in plain language.

This multimodal LLM also annotates images in the dataset by identifying which concepts are present and absent in each image. The researchers use this annotated dataset to train a concept bottleneck module to recognize the concepts.

They incorporate this module into the target model, forcing it to make predictions using only the set of learned concepts the researchers extracted.

Controlling the concepts

They overcame many challenges as they developed this method, from ensuring the LLM annotated concepts correctly to determining whether the sparse autoencoder had identified human-understandable concepts.

To prevent the model from using unknown or unwanted concepts, they restrict it to use only five concepts for each prediction. This also forces the model to choose the most relevant concepts and makes the explanations more understandable.

When they compared their approach to state-of-the-art CBMs on tasks like predicting bird species and identifying skin lesions in medical images, their method achieved the highest accuracy while providing more precise explanations.

Their approach also generated concepts that were more applicable to the images in the dataset.

“We’ve shown that extracting concepts from the original model can outperform other CBMs, but there is still a tradeoff between interpretability and accuracy that needs to be addressed. Black-box models that are not interpretable still outperform ours,” De Santis says.

In the future, the researchers want to study potential solutions to the information leakage problem, perhaps by adding additional concept bottleneck modules so unwanted concepts can’t leak through. They also plan to scale up their method by using a larger multimodal LLM to annotate a bigger training dataset, which could boost performance.

“I’m excited by this work because it pushes interpretable AI in a very promising direction and creates a natural bridge to symbolic AI and knowledge graphs,” says Andreas Hotho, professor and head of the Data Science Chair at the University of Würzburg, who was not involved with this work. “By deriving concept bottlenecks from the model’s own internal mechanisms rather than only from human-defined concepts, it offers a path toward explanations that are more faithful to the model and opens many opportunities for follow-up work with structured knowledge.”

This research was supported by the Progetto Rocca Doctoral Fellowship, the Italian Ministry of University and Research under the National Recovery and Resilience Plan, Thales Alenia Space, and the European Union under the NextGenerationEU project.

The post Improving AI models’ ability to explain their predictions appeared first on MIT Schwarzman College of Computing.

Consider car safety design. Engineers must integrate thousands of parts, and many design choices can affect how a vehicle performs in a collision. Classic optimization tools could start to struggle when searching for the best combination.

MIT researchers developed a new approach that rethinks how a classic method, known as Bayesian optimization, can be used to solve problems with hundreds of variables. In tests on realistic engineering-style benchmarks, like power-system optimization, the approach found top solutions 10 to 100 times faster than widely used methods.

Their technique leverages a foundation model trained on tabular data that automatically identifies the variables that matter most for improving performance, repeating the process to hone in on better and better solutions. Foundation models are huge artificial intelligence systems trained on vast, general datasets. This allows them to adapt to different applications.

The researchers’ tabular foundation model does not need to be constantly retrained as it works toward a solution, increasing the efficiency of the optimization process. The technique also delivers greater speedups for more complicated problems, so it could be especially useful in demanding applications like materials development or drug discovery.

“Modern AI and machine-learning models can fundamentally change the way engineers and scientists create complex systems. We came up with one algorithm that can not only solve high-dimensional problems, but is also reusable so it can be applied to many problems without the need to start everything from scratch,” says Rosen Yu, a graduate student in computational science and engineering and lead author of a paper on this technique.

Yu is joined on the paper by Cyril Picard, a former MIT postdoc and research scientist, and Faez Ahmed, associate professor of mechanical engineering and a core member of the MIT Center for Computational Science and Engineering. The research will be presented at the International Conference on Learning Representations.

Improving a proven method

When scientists seek to solve a multifaceted problem but have expensive methods to evaluate success, like crash testing a car to know how good each design is, they often use a tried-and-true method called Bayesian optimization. This iterative method finds the best configuration for a complicated system by building a surrogate model that helps estimate what to explore next while considering the uncertainty of its predictions.

But the surrogate model must be retrained after each iteration, which can quickly become computationally intractable when the space of potential solutions is very large. In addition, scientists need to build a new model from scratch any time they want to tackle a different scenario.

To address both shortcomings, the MIT researchers utilized a generative AI system known as a tabular foundation model as the surrogate model inside a Bayesian optimization algorithm.

“A tabular foundation model is like a ChatGPT for spreadsheets. The input and output of these models are tabular data, which in the engineering domain is much more common to see and use than language,” Yu says.

Just like large language models such as ChatGPT,  Claude, and Gemini, the model has been pre-trained on an enormous amount of tabular data. This makes it well-equipped to tackle a range of prediction problems. In addition, the model can be deployed as-is, without the need for any retraining.

To make their system more accurate and efficient for optimization, the researchers employed a trick that enables the model to identify features of the design space that will have the biggest impact on the solution.

“A car might have 300 design criteria, but not all of them are the main driver of the best design if you are trying to increase some safety parameters. Our algorithm can smartly select the most critical features to focus on,” Yu says.

It does this by using a tabular foundation model to estimate which variables (or combinations of variables) most influence the outcome.

It then focuses the search on those high-impact variables instead of wasting time exploring everything equally. For instance, if the size of the front crumple zone significantly increased and the car’s safety rating improved, that feature likely played a role in the enhancement.

Bigger problems, better solutions

One of their biggest challenges was finding the best tabular foundation model for this task, Yu says. Then they had to connect it with a Bayesian optimization algorithm in such a way that it could identify the most prominent design features.

“Finding the most prominent dimension is a well-known problem in math and computer science, but coming up with a way that leveraged the properties of a tabular foundation model was a real challenge,” Yu says.

With the algorithmic framework in place, the researchers tested their method by comparing it to five state-of-the-art optimization algorithms.

On 60 benchmark problems, including realistic situations like power grid design and car crash testing, their method consistently found the best solution between 10 and 100 times faster than the other algorithms.

“When an optimization problem gets more and more dimensions, our algorithm really shines,” Yu added.

But their method did not outperform the baselines on all problems, such as robotic path planning. This likely indicates that scenario was not well-defined in the model’s training data, Yu says.

In the future, the researchers want to study methods that could boost the performance of tabular foundation models. They also want to apply their technique to problems with thousands or even millions of dimensions, like the design of a naval ship.

“At a higher level, this work points to a broader shift: using foundation models not just for perception or language, but as algorithmic engines inside scientific and engineering tools, allowing classical methods like Bayesian optimization to scale to regimes that were previously impractical,” says Ahmed.

“The approach presented in this work, using a pretrained foundation model together with high‑dimensional Bayesian optimization, is a creative and promising way to reduce the heavy data requirements of simulation‑based design. Overall, this work is a practical and powerful step toward making advanced design optimization more accessible and easier to apply in real-world settings,” says Wei Chen, the Wilson-Cook Professor in Engineering Design and chair of the Department of Mechanical Engineering at Northwestern University, who was not involved in this research.

The post A “ChatGPT for spreadsheets” helps solve difficult engineering challenges faster appeared first on MIT Schwarzman College of Computing.

But developing reasoning models demands an enormous amount of computation and energy due to inefficiencies in the training process. While a few of the high-power processors continuously work through complicated queries, others in the group sit idle.

Researchers from MIT and elsewhere found a way to use this computational downtime to efficiently accelerate reasoning-model training.

Their new method automatically trains a smaller, faster model to predict the outputs of the larger reasoning LLM, which the larger model verifies. This reduces the amount of work the reasoning model must do, accelerating the training process.

The key to this system is its ability to train and deploy the smaller model adaptively, so it kicks in only when some processors are idle. By leveraging computational resources that would otherwise have been wasted, it accelerates training without incurring additional overhead.

When tested on multiple reasoning LLMs, the method doubled the training speed while preserving accuracy. This could reduce the cost and increase the energy efficiency of developing advanced LLMs for applications such as forecasting financial trends or detecting risks in power grids.

“People want models that can handle more complex tasks. But if that is the goal of model development, then we need to prioritize efficiency. We found a lossless solution to this problem and then developed a full-stack system that can deliver quite dramatic speedups in practice,” says Qinghao Hu, an MIT postdoc and co-lead author of a paper on this technique.

He is joined on the paper by co-lead author Shang Yang, an electrical engineering and computer science (EECS) graduate student; Junxian Guo, an EECS graduate student; senior author Song Han, an associate professor in EECS, member of the Research Laboratory of Electronics and a distinguished scientist of NVIDIA; as well as others at NVIDIA, ETH Zurich, the MIT-IBM Watson AI Lab, and the University of Massachusetts at Amherst. The research will be presented at the ACM International Conference on Architectural Support for Programming Languages and Operating Systems.

Training bottleneck

Developers want reasoning LLMs to identify and correct mistakes in their critical thinking process. This capability allows them to ace complicated queries that would trip up a standard LLM.

To teach them this skill, developers train reasoning LLMs using a technique called reinforcement learning (RL). The model generates multiple potential answers to a query, receives a reward for the best candidate, and is updated based on the top answer. These steps repeat thousands of times as the model learns.

But the researchers found that the process of generating multiple answers, called rollout, can consume as much as 85 percent of the execution time needed for RL training.

“Updating the model — which is the actual ‘training’ part — consumes very little time by comparison,” Hu says.

This bottleneck occurs in standard RL algorithms because all processors in the training group must finish their responses before they can move on to the next step. Because some processors might be working on very long responses, others that generated shorter responses wait for them to finish.

“Our goal was to turn this idle time into speedup without any wasted costs,” Hu adds.

They sought to use an existing technique, called speculative decoding, to speed things up. Speculative decoding involves training a smaller model called a drafter to rapidly guess the future outputs of the larger model.

The larger model verifies the drafter’s guesses, and the responses it accepts are used for training.

Because the larger model can verify all the drafter’s guesses at once, rather than generating each output sequentially, it accelerates the process.

An adaptive solution

But in speculative decoding, the drafter model is typically trained only once and remains static. This makes the technique infeasible for reinforcement learning, since the reasoning model is updated thousands of times during training.

A static drafter would quickly become stale and useless after a few steps.

To overcome this problem, the researchers created a flexible system known as “Taming the Long Tail,” or TLT.

The first part of TLT is an adaptive drafter trainer, which uses free time on idle processors to train the drafter model on the fly, keeping it well-aligned with the target model without using extra computational resources.

The second component, an adaptive rollout engine, manages speculative decoding to automatically select the optimal strategy for each new batch of inputs. This mechanism changes the speculative decoding configuration based on the training workload features, such as the number of inputs processed by the draft model and the number of inputs accepted by the target model during verification.

In addition, the researchers designed the draft model to be lightweight so it can be trained quickly. TLT reuses some components of the reasoning model training process to train the drafter, leading to extra gains in acceleration.

“As soon as some processors finish their short queries and become idle, we immediately switch them to do draft model training using the same data they are using for the rollout process. The key mechanism is our adaptive speculative decoding — these gains wouldn’t be possible without it,” Hu says.

They tested TLT across multiple reasoning LLMs that were trained using real-world datasets. The system accelerated training between 70 and 210 percent while preserving the accuracy of each model.

As an added bonus, the small drafter model could readily be utilized for efficient deployment as a free byproduct.

In the future, the researchers want to integrate TLT into more types of training and inference frameworks and find new reinforcement learning applications that could be accelerated using this approach.

“As reasoning continues to become the major workload driving the demand for inference, Qinghao’s TLT is great work to cope with the computation bottleneck of training these reasoning models. I think this method will be very helpful in the context of efficient AI computing,” Han says.

This work is funded by the MIT-IBM Watson AI Lab, the MIT AI Hardware Program, the MIT Amazon Science Hub, Hyundai Motor Company, and the National Science Foundation.

The post New method could increase LLM training efficiency appeared first on MIT Schwarzman College of Computing.

“My current research focuses on applying AI techniques to cybersecurity problems and examining the policy implications of these advancements, especially in the context of maritime cybersecurity,” says Janjusevic. “Representing my country at the highest levels of education and industry has given me a unique perspective on cybersecurity challenges.”

Janjusevic’s pathway from Montenegro to Maryland was created by a program that allows selected students from allied countries to attend the U.S. Naval Academy. Janjusevic graduated with a dual bachelor’s degree in cyber operations and computer science. His undergraduate experience provided opportunities to collaborate with the U.S. military and the National Security Agency, exposing him to high-level cybersecurity operations and fueling his interest in tackling complex cybersecurity challenges. During his undergraduate studies, he also interned with Microsoft, developing tools for cloud incident response, and with NASA, visualizing solar data for research applications.

Following his graduation, he realized that he still needed more knowledge, particularly in the area of AI and cybersecurity. TPP appealed to him immediately because of its dual emphasis on rigorous engineering innovation and the policy analysis needed to deploy it effectively. Janjusevic’s experiences at TPP have been a big change from his time at the U.S. Naval Academy, with a different pace and environment. He has especially appreciated being able to broaden his understanding about a variety of research domains and apply the discipline and knowledge he earned during his time at the academy.

“My TPP experience has been amazing,” says Janjusevic. “The cohort is really small, so it feels like a family, and everyone is working on diverse, high-impact problems.”

Mitigating the risks of emerging technologies

Janjusevic’s thesis brings together disciplines of cybersecurity, AI and deep learning, and control theory and physics, focusing on securing maritime cyber-physical systems — in particular, large legacy ships. The hacking of these ships’ networks can result in substantial damage to national security, as well as serious economic effects.

“Strajo is working to outsmart maritime GPS spoofing,” says Saurabh Amin, the Edmund K. Turner Professor in Civil Engineering. “Such attacks have already lured vessels off course in contested waters. His approach layers physics-based trajectory models with deep learning, catching threats that no single method can detect alone. His expertise has been very helpful in advancing our work on threat modeling and attack detection.”

The research utilizes advanced threat modeling and vessel dynamics to train AI systems to distinguish between legitimate maneuvers and spoofed signals. It involves building a framework that employs an internal LSTM (long short-term memory) autoencoder to analyze signal integrity, while simultaneously using a physics-based forecaster to predict the vessel’s movement based on environmental factors like wind and the sea state. By comparing these predictions against reported GPS positions, the system can effectively distinguish between natural sensor noise and malicious spoofing attacks. This hybrid framework is designed to empower, not replace, human operators, providing verified navigation data that allows watch standers to distinguish technical glitches from strategic cyberattacks.

Janjusevic has been able to enhance his academic research with industry experience. In summer 2025, he interned with the Network Detection team at the AI cybersecurity company Vectra AI. There, he investigated potential threats new technologies can bring, particularly AI agents and the model context protocol (MCP) — the emerging standard for AI agent communication. His research demonstrated how this technology could be repurposed for autonomous hacking operations and advanced command and control. This work on the security risks of agentic AI was recently presented in the preprint, “Hiding in the AI Traffic: Abusing MCP for LLM-Powered Agentic Red Teaming.”

“I was able to gain practical insights and hands-on experience into how a data science team uses AI models to detect anomalies in a network,” says Janjusevic. “This work within industry directly informed the anomaly detection models in my research.”

International policy perspective

“Strajo brings not just a high level of intelligence and energy to his work on cyber-physical security for merchant vessels, but also a strong instinct from his Navy training that resonates deeply with the research effort and grounds it in actionable policy,” says Fotini Christia, the Ford International Professor of the Social Sciences, director of IDSS, and a leader of the MIT Maritime Consortium.

Janjusevic participates in the cybersecurity efforts of the Maritime Consortium, a collaboration between academia, industry, and regulatory agencies focused on developing technological solutions, industry standards, and policies. The consortium includes cooperation with some international members, including from Singapore and South Korea.

“In AI cybersecurity, the policy element is really important, as the field is so fast-moving and the consequences of hacking can be so dangerous,” says Janjusevic. “I think there’s still a lot of need for policy work in this space.”

Janjusevic is also currently helping to organize two upcoming major conferences: the Harvard European Conference in February, which will convene officials and diplomats from across the globe, and the Technology and National Security Conference in April, a collaboration of Harvard and MIT that brings together top leaders from government, industry, and academia to tackle critical challenges in national security.

“I’m striving to find a position where I can influence and advance the cybersecurity field with AI, while at the same time leading collaboration and innovation between the United States and Montenegro,” says Janjusevic. “My goal is to be a bridge between Europe and the U.S. in this space of national security, AI, and cybersecurity, bringing my knowledge to both sides.”

The post Enhancing maritime cybersecurity with technology and policy appeared first on MIT Schwarzman College of Computing.

Where in the cell the information comes from matters. But to capture complete information about the state of the cell, scientists often must conduct many measurements using different techniques and analyze them one at a time. Machine-learning methods can speed up the process, but existing methods lump all the information from each measurement modality together, making it difficult to figure out which data came from which part of the cell.

To overcome this problem, researchers at the Broad Institute of MIT and Harvard and ETH Zurich/Paul Scherrer Institute (PSI) developed an artificial intelligence-driven framework that learns which information about a cell’s state is shared across different measurement modalities and which information is unique to a particular measurement type.

By pinpointing which information came from which cell parts, the approach provides a more holistic view of the cell’s state, making it easier for a biologist to see the complete picture of cellular interactions. This could help scientists understand disease mechanisms and track the progression of cancer, neurodegenerative disorders such as Alzheimer’s, and metabolic diseases like diabetes.

“When we study cells, one measurement is often not sufficient, so scientists develop new technologies to measure different aspects of cells. While we have many ways of looking at a cell, at the end of the day we only have one underlying cell state. By putting the information from all these measurement modalities together in a smarter way, we could have a fuller picture of the state of the cell,” says lead author Xinyi Zhang SM ’22, PhD ’25, a former graduate student in the MIT Department of Electrical Engineering and Computer Science (EECS) and an affiliate of the Eric and Wendy Schmidt Center at the Broad Institute of MIT and Harvard, who is now a group leader at AITHYRA in Vienna, Austria.

Zhang is joined on a paper about the work by G.V. Shivashankar, a professor in the Department of Health Sciences and Technology at ETH Zurich and head of the Laboratory of Multiscale Bioimaging at PSI; and senior author Caroline Uhler, a professor in EECS and the Institute for Data, Systems, and Society (IDSS) at MIT, member of MIT’s Laboratory for Information and Decision Systems (LIDS), and director of the Eric and Wendy Schmidt Center at the Broad Institute. The research appears today in Nature Computational Science.

Manipulating multiple measurements

There are many tools scientists can use to capture information about a cell’s state. For instance, they can measure RNA to see if the cell is growing, or they can measure chromatin morphology to see if the cell is dealing with external physical or chemical signals.

“When scientists perform multimodal analysis, they gather information using multiple measurement modalities and integrate it to better understand the underlying state of the cell. Some information is captured by one modality only, while other information is shared across modalities. To fully understand what is happening inside the cell, it is important to know where the information came from,” says Shivashankar.

Often, for scientists, the only way to sort this out is to conduct multiple individual experiments and compare the results. This slow and cumbersome process limits the amount of information they can gather.

In the new work, the researchers built a machine-learning framework that specifically understands which information overlaps between different modalities, and which information is unique to a particular modality but not captured by others.

“As a user, you can simply input your cell data and it automatically tells you which data are shared and which data are modality-specific,” Zhang says.

To build this framework, the researchers rethought the typical way machine-learning models are designed to capture and interpret multimodal cellular measurements.

Usually these methods, known as autoencoders, have one model for each measurement modality, and each model encodes a separate representation for the data captured by that modality. The representation is a compressed version of the input data that discards any irrelevant details.

The MIT method has a shared representation space where data that overlap between multiple modalities are encoded, as well as separate spaces where unique data from each modality are encoded.

In essence, one can think of it like a Venn diagram of cellular data.

The researchers also used a special, two-step training procedure that helps their model handle the complexity involved in deciding which data are shared across multiple data modalities. After training, the model can identify which data are shared and which are unique when fed cell data it has never seen before.

Distinguishing data

In tests on synthetic datasets, the framework correctly captured known shared and modality-specific information. When they applied their method to real-world single-cell datasets, it comprehensively and automatically distinguished between gene activity captured jointly by two measurement modalities, such as transcriptomics and chromatin accessibility, while also correctly identifying which information came from only one of those modalities.

In addition, the researchers used their method to identify which measurement modality captured a certain protein marker that indicates DNA damage in cancer patients. Knowing where this information came from would help a clinical scientist determine which technique they should use to measure that marker.

“There are too many modalities in a cell and we can’t possibly measure them all, so we need a prediction tool. But then the question is: Which modalities should we measure and which modalities should we predict? Our method can answer that question,” Uhler says.

In the future, the researchers want to enable the model to provide more interpretable information about the state of the cell. They also want to conduct additional experiments to ensure it correctly disentangles cellular information and apply the model to a wider range of clinical questions.

“It is not sufficient to just integrate the information from all these modalities,” Uhler says. “We can learn a lot about the state of a cell if we carefully compare the different modalities to understand how different components of cells regulate each other.”

This research is funded, in part, by the Eric and Wendy Schmidt Center at the Broad Institute, the Swiss National Science Foundation, the U.S. National Institutes of Health, the U.S. Office of Naval Research, AstraZeneca, the MIT-IBM Watson AI Lab, the MIT J-Clinic for Machine Learning and Health, and a Simons Investigator Award.

The post AI to help researchers see the bigger picture in cell biology appeared first on MIT Schwarzman College of Computing.