A causal theory for studying the cause-and-effect relationships of genes

Posted on November 12, 2024 by - News

By studying changes in gene expression, researchers learn how cells function at a molecular level, which could help them understand the development of certain diseases.

But a human has about 20,000 genes that can affect each other in complex ways, so even knowing which groups of genes to target is an enormously complicated problem. Also, genes work together in modules that regulate each other.

MIT researchers have now developed theoretical foundations for methods that could identify the best way to aggregate genes into related groups so they can efficiently learn the underlying cause-and-effect relationships between many genes.

Importantly, this new method accomplishes this using only observational data. This means researchers don’t need to perform costly, and sometimes infeasible, interventional experiments to obtain the data needed to infer the underlying causal relationships.

In the long run, this technique could help scientists identify potential gene targets to induce certain behavior in a more accurate and efficient manner, potentially enabling them to develop precise treatments for patients.

“In genomics, it is very important to understand the mechanism underlying cell states. But cells have a multiscale structure, so the level of summarization is very important, too. If you figure out the right way to aggregate the observed data, the information you learn about the system should be more interpretable and useful,” says graduate student Jiaqi Zhang, an Eric and Wendy Schmidt Center Fellow and co-lead author of a paper on this technique.

Zhang is joined on the paper by co-lead author Ryan Welch, currently a master’s student in engineering; and senior author Caroline Uhler, a professor in the Department of Electrical Engineering and Computer Science (EECS) and the Institute for Data, Systems, and Society (IDSS) who is also director of the Eric and Wendy Schmidt Center at the Broad Institute of MIT and Harvard, and a researcher at MIT’s Laboratory for Information and Decision Systems (LIDS). The research will be presented at the Conference on Neural Information Processing Systems.

Learning from observational data

The problem the researchers set out to tackle involves learning programs of genes. These programs describe which genes function together to regulate other genes in a biological process, such as cell development or differentiation.

Since scientists can’t efficiently study how all 20,000 genes interact, they use a technique called causal disentanglement to learn how to combine related groups of genes into a representation that allows them to efficiently explore cause-and-effect relationships.

In previous work, the researchers demonstrated how this could be done effectively in the presence of interventional data, which are data obtained by perturbing variables in the network.

But it is often expensive to conduct interventional experiments, and there are some scenarios where such experiments are either unethical or the technology is not good enough for the intervention to succeed.

With only observational data, researchers can’t compare genes before and after an intervention to learn how groups of genes function together.

“Most research in causal disentanglement assumes access to interventions, so it was unclear how much information you can disentangle with just observational data,” Zhang says.

The MIT researchers developed a more general approach that uses a machine-learning algorithm to effectively identify and aggregate groups of observed variables, e.g., genes, using only observational data.

They can use this technique to identify causal modules and reconstruct an accurate underlying representation of the cause-and-effect mechanism. “While this research was motivated by the problem of elucidating cellular programs, we first had to develop novel causal theory to understand what could and could not be learned from observational data. With this theory in hand, in future work we can apply our understanding to genetic data and identify gene modules as well as their regulatory relationships,” Uhler says.

A layerwise representation

Using statistical techniques, the researchers can compute a mathematical function known as the variance for the Jacobian of each variable’s score. Causal variables that don’t affect any subsequent variables should have a variance of zero.

The researchers reconstruct the representation in a layer-by-layer structure, starting by removing the variables in the bottom layer that have a variance of zero. Then they work backward, layer-by-layer, removing the variables with zero variance to determine which variables, or groups of genes, are connected.

“Identifying the variances that are zero quickly becomes a combinatorial objective that is pretty hard to solve, so deriving an efficient algorithm that could solve it was a major challenge,” Zhang says.

In the end, their method outputs an abstracted representation of the observed data with layers of interconnected variables that accurately summarizes the underlying cause-and-effect structure.

Each variable represents an aggregated group of genes that function together, and the relationship between two variables represents how one group of genes regulates another. Their method effectively captures all the information used in determining each layer of variables.

After proving that their technique was theoretically sound, the researchers conducted simulations to show that the algorithm can efficiently disentangle meaningful causal representations using only observational data.

In the future, the researchers want to apply this technique in real-world genetics applications. They also want to explore how their method could provide additional insights in situations where some interventional data are available, or help scientists understand how to design effective genetic interventions. In the future, this method could help researchers more efficiently determine which genes function together in the same program, which could help identify drugs that could target those genes to treat certain diseases.

This research is funded, in part, by the U.S. Office of Naval Research, the National Institutes of Health, the U.S. Department of Energy, a Simons Investigator Award, the Eric and Wendy Schmidt Center at the Broad Institute, the Advanced Undergraduate Research Opportunities Program at MIT, and an Apple AI/ML PhD Fellowship.

3 questions: Leveraging insights to enable clinical outcomes

Posted on November 12, 2024 by - News

Associate Professor Thomas Heldt joined the MIT faculty in 2013 as a core member of the Institute for Medical Engineering and Science (IMES) and the Department of Electrical Engineering and Computer Science. Additionally, Heldt is a principal investigator with MIT’s Research Laboratory of Electronics (RLE), and he directs the Integrative Neuromonitoring and Critical Care Informatics Group in IMES and RLE. He was recently named an associate director of IMES, where he will focus on internal affairs, among other duties. 

Heldt received his Medical Engineering and Medical Physics (MEMP) PhD from the Harvard-MIT Program in Health Sciences and Technology (HST) in 2004. Heldt’s research interests include signal processing, estimation and identification of physiological systems, mathematical modeling, model identification to support real-time clinical decision making, monitoring of disease progression, and titration of therapy, primarily in neurocritical and neonatal critical care. Here, Heldt describes how he collaborates closely with MIT colleagues and others at Boston-area hospitals, and how his research uses and analyzes physiologic data to aid clinical action.

Q: How does your research apply to solving clinical needs?

A: We look at current clinical environments and observe the volumes of multimodal physiologic waveform data that are collected on patients in critical care, peri-operative care, or even emergency care. Much of this data is typically visually reviewed by the clinicians and subsequently discarded after a holding period of just a few days. We thus lose the opportunity for more systematic analyses and for deriving patient-specific insights. Critical to such analyses of these data streams is a deep understanding of the relevant physiology at the time scales of interest. We leverage insights from physiology, formulated as reduced order mathematical models capturing the essential mechanisms that enable clinical action. We have applied this approach successfully to estimate intracranial pressure noninvasively, to make diagnostic decisions based on the analysis of the shape of the capnogram, and, are currently using ultrasound-based approaches to detect embolic events in patients on life support, such as ventricular assist devices or extracorporeal membrane oxygenation. 

Q: You work closely with colleagues across MIT, and with clinicians at Boston-area hospitals, including Boston Children’s Hospital (where you hold a courtesy research appointment in neurology), Boston Medical Center (neurosurgery), and Massachusetts General Hospital (emergency medicine). What has been the fruit of some of these collaborations — what is the impact on your research?

A: Boston is a fantastic place to conduct translational research that crosses from our laboratories at MIT into the clinical environments for validation in the actual target patient population! The collaborative disposition and forward-thinking mindset of our clinician colleagues have really been fundamentally enabling for our research and have provided amazing mentoring to our students, postdocs, and me. We have collected validation data in brain-injured patients in the ICUs [intensive care units] at Boston Medical Center, Boston Children’s Hospital (BCH), and Beth Israel Deaconess Medical Center (BIDMC); we have collected pilot and validation data for our capnography work in the emergency departments at BCH and BIDMC; we have collected data for our emboli work in the operating rooms and ICUs at BCH, and have analyzed the medical records of the neonatal ICU at BIDMC and the emergency department at Massachusetts General Hospital.

Our work with the neonatologist at BIDMC was focused on analyzing the monitoring alarm patterns in the neonatal ICU. We counted a staggering 177 alarms/baby/day, or one alarm every eight minutes on average, per baby. And this is a 54-bed neonatal ICU operating close to capacity every day! Such volumes of alarms contribute to noise pollution in an environment that should ideally be very calm. Additionally, since most of the alarms are nuisance alarms or do not require any clinical intervention, the clinical staff becomes desensitized to the alarm load and might end up ignoring truly important events. We analyzed the alarm patterns and alarm thresholds for a particular type of heart rate alarms and recommended a change in thresholds. This resulted in a 50 percent reduction in heart rate alarms per patient per day. Initially, the clinical staff had to file weekly reports to make sure the reduction in the alarm rate did not result in missed or adverse events. After about three months without a single reportable event, the hospital safety committee approved the change.

With colleagues from the MGH Department of Emergency Medicine, we developed and tested a triage rule to identify patients at risk of septic shock. At the time, the MGH ED [emergency department] saw more than 120,000 patients/year, and around 75 percent of patients ending up in the ICU with severe sepsis and septic shock came through the emergency department. Hence, ED triage was the first point of patient contact and the first opportunity to flag patients for possible sepsis and septic shock and initiation of early goal-directed therapy. One result of our work was a significant reduction in the time to appropriate antibiotic administration in the emergency department. The work was subsequently validated in other Partners hospitals and implemented in the electronic medical record system of Partners-affiliated hospitals. 

Q: Can you talk a bit about your background, and about how you became interested in systems-physiology and biomedicine? What are your goals for your research, and for your career?

A: That is a longer story! In short, I started out studying physics back in Germany. After a while, I got interested in applying concepts I learned in physics to physiology and medicine, so I designed my own MD/PhD program by picking up medicine as a second major. Through some fortuitous events, I ended up attending surgeries for congenital heart defects for about a term. This was a very formative experience, and almost pushed me toward dropping physics and going all-out on becoming a surgeon. However, I had also always wanted to spend part of my education abroad and had applied to various universities in the U.S. I ended up getting admitted to the graduate physics program at Yale and spent a couple of years doing nonlinear optics. While I loved the work at Yale and had a fantastic mentor, I missed the clinical exposure and application of my work to medicine. I had heard about the HST program and decided to send in an application. I joined the MEMP program in 1997 and have been at MIT ever since.

In our current research, we are very interested in providing better monitoring modalities for patients with brain injuries. We are developing novel algorithmic and device approaches so we can replace the current invasive monitoring modalities with entirely noninvasive ones and provide additional clinically actionable information that gives insights on the physiology of the injured brain and can help guide treatment decision. I want to see some of these technologies through to routine deployment at the bedside.

The great thing about being in IMES and MIT is that everybody is very collaborative. What I am looking forward to is much of the same, working with colleagues in IMES on important problems that none of us is be able to tackle alone, but that together we have a real chance of tackling — and having fun along the way! 

Empowering systemic racism research at MIT and beyond

Posted on November 6, 2024 by - News

Caption:The new ICSR Data Hub serves as an evolving, public web depository of datasets gathered by MIT researchers examining racial bias in American society and institutions.

At the turn of the 20th century, W.E.B. Du Bois wrote about the conditions and culture of Black people in Philadelphia, documenting also the racist attitudes and beliefs that pervaded the white society around them. He described how unequal outcomes in domains like health could be attributed not only to racist ideas, but to racism embedded in American institutions.

Almost 125 years later, the concept of “systemic racism” is central to the study of race. Centuries of data collection and analysis, like the work of Du Bois, document the mechanisms of racial inequity in law and institutions, and attempt to measure their impact.

“There’s extensive research showing racial discrimination and systemic inequity in essentially all sectors of American society,” explains Fotini Christia, the Ford International Professor of Social Sciences in the Department of Political Science, who directs the MIT Institute for Data, Systems, and Society (IDSS), where she also co-leads the Initiative on Combatting Systemic Racism (ICSR). “Newer research demonstrates how computational technologies, typically trained or reliant on historical data, can further entrench racial bias. But these same tools can also help to identify racially inequitable outcomes, to understand their causes and impacts, and even contribute to proposing solutions.”

In addition to coordinating research on systemic racism across campus, the IDSS initiative has a new project aiming to empower and support this research beyond MIT: the new ICSR Data Hub, which serves as an evolving, public web depository of datasets gathered by ICSR researchers.

Data for justice

“My main project with ICSR involved using Amazon Web Services to build the data hub for other researchers to use in their own criminal justice related projects,” says Ben Lewis SM ’24, a recent alumnus of the MIT Technology and Policy Program (TPP) and current doctoral student at the MIT Sloan School of Management. “We want the data hub to be a centralized place where researchers can access this information via a simple web or Python interface.”

While earning his master’s degree at TPP, Lewis focused his research on race, drug policy, and policing in the United States, exploring drug decriminalization policies’ impact on rates of incarceration and overdose. He worked as a member of the ICSR Policing team, a group of researchers across MIT examining the roles data plays in the design of policing policies and procedures, and how data can highlight or exacerbate racial bias.

“The Policing vertical started with a really challenging fundamental question,” says team lead and electrical engineering and computer science (EECS) Professor Devavrat Shah. “Can we use data to better understand the role that race plays in the different decisions made throughout the criminal justice system?”

So far, the data hub offers 911 dispatch information and police stop data, gathered from 40 of the largest cities in the United States by ICSR researchers. Lewis hopes to see the effort expand to include not only other cities, but other relevant and typically siloed information, like sentencing data.

“We want to stitch the datasets together so that we have a more comprehensive and holistic view of law enforcement systems,” explains Jessy Xinyi Han, a fellow ICSR researcher and graduate student in the IDSS Social and Engineering Systems (SES) doctoral program. Statistical methods like causal inference can help to uncover root causes behind inequalities, says Han — to “untangle a web of possibilities” and better understand the causal effect of race at different stages of the criminal justice process.

“My motivation behind doing this project is personal,” says Lewis, who was drawn to MIT in large part by the opportunity to research systemic racism. As a TPP student, he also founded the Cambridge branch of End Overdose, a nonprofit dedicated to stopping drug overdose deaths. His advocacy led to training hundreds in lifesaving drug interventions, and earned him the 2024 Collier Medal, an MIT distinction for community service honoring Sean Collier, who gave his life serving as an officer with the MIT Police.

“I’ve had family members in incarceration. I’ve seen the impact it has had on my family, and on my community, and realized that over-policing and incarceration are a Band-Aid on issues like poverty and drug use that can trap people in a cycle of poverty.”

Education and impact

Now that the infrastructure for the data hub has been built, and the ICSR Policing team has begun sharing datasets, the next step is for other ICSR teams to start sharing data as well. The cross-disciplinary systemic racism research initiative includes teams working in domains including housing, health care, and social media.

“We want to take advantage of the abundance of data that is available today to answer difficult questions about how racism results from the interactions of multiple systems,” says Munther Dahleh, EECS professor, IDSS founding director, and ICSR co-lead. “Our interest is in how various institutions perpetuate racism, and how technology can exacerbate or combat this.”

To the data hub creators, the main sign of success for the project is seeing the data used in research projects at and beyond MIT. As a resource, though, the hub can support that research for users from a range of experience and backgrounds.

“The data hub is also about education and empowerment,” says Han. “This information can be used in projects designed to teach users how to use big data, how to do data analysis, and even to learn machine learning tools, all specifically to uncover racial disparities in data.”

“Championing the propagation of data skills has been part of the IDSS mission since Day 1,” says Dahleh. “We are excited by the opportunities that making this data available can present in educational contexts, including but not limited to our growing IDSSx suite of online course offerings.”

This emphasis on educational potential only augments the ambitions of ICSR researchers across MIT, who aspire to use data and computing tools to produce actionable insights for policymakers that can lead to real change.

“Systemic racism is an abundantly evidenced societal challenge with far-reaching impacts across domains,” says Christia. “At IDSS, we want to ensure that developing technologies, combined with access to ever-increasing amounts of data, are leveraged to combat racist outcomes rather than continue to enact them.”

Despite its impressive output, generative AI doesn’t have a coherent understanding of the world

Posted on November 6, 2024 by - News

Such surprising capabilities can make it seem like the models are implicitly learning some general truths about the world.

But that isn’t necessarily the case, according to a new study. The researchers found that a popular type of generative AI model can provide turn-by-turn driving directions in New York City with near-perfect accuracy — without having formed an accurate internal map of the city.

Despite the model’s uncanny ability to navigate effectively, when the researchers closed some streets and added detours, its performance plummeted.

When they dug deeper, the researchers found that the New York maps the model implicitly generated had many nonexistent streets curving between the grid and connecting far away intersections.

This could have serious implications for generative AI models deployed in the real world, since a model that seems to be performing well in one context might break down if the task or environment slightly changes.

“One hope is that, because LLMs can accomplish all these amazing things in language, maybe we could use these same tools in other parts of science, as well. But the question of whether LLMs are learning coherent world models is very important if we want to use these techniques to make new discoveries,” says senior author Ashesh Rambachan, assistant professor of economics and a principal investigator in the MIT Laboratory for Information and Decision Systems (LIDS).

Rambachan is joined on a paper about the work by lead author Keyon Vafa, a postdoc at Harvard University; Justin Y. Chen, an electrical engineering and computer science (EECS) graduate student at MIT; Jon Kleinberg, Tisch University Professor of Computer Science and Information Science at Cornell University; and Sendhil Mullainathan, an MIT professor in the departments of EECS and of Economics, and a member of LIDS. The research will be presented at the Conference on Neural Information Processing Systems.

New metrics

The researchers focused on a type of generative AI model known as a transformer, which forms the backbone of LLMs like GPT-4. Transformers are trained on a massive amount of language-based data to predict the next token in a sequence, such as the next word in a sentence.

But if scientists want to determine whether an LLM has formed an accurate model of the world, measuring the accuracy of its predictions doesn’t go far enough, the researchers say.

For example, they found that a transformer can predict valid moves in a game of Connect 4 nearly every time without understanding any of the rules.

So, the team developed two new metrics that can test a transformer’s world model. The researchers focused their evaluations on a class of problems called deterministic finite automations, or DFAs. 

A DFA is a problem with a sequence of states, like intersections one must traverse to reach a destination, and a concrete way of describing the rules one must follow along the way.

They chose two problems to formulate as DFAs: navigating on streets in New York City and playing the board game Othello.

“We needed test beds where we know what the world model is. Now, we can rigorously think about what it means to recover that world model,” Vafa explains.

The first metric they developed, called sequence distinction, says a model has formed a coherent world model it if sees two different states, like two different Othello boards, and recognizes how they are different. Sequences, that is, ordered lists of data points, are what transformers use to generate outputs.

The second metric, called sequence compression, says a transformer with a coherent world model should know that two identical states, like two identical Othello boards, have the same sequence of possible next steps.

They used these metrics to test two common classes of transformers, one which is trained on data generated from randomly produced sequences and the other on data generated by following strategies.

Incoherent world models

Surprisingly, the researchers found that transformers which made choices randomly formed more accurate world models, perhaps because they saw a wider variety of potential next steps during training. 

“In Othello, if you see two random computers playing rather than championship players, in theory you’d see the full set of possible moves, even the bad moves championship players wouldn’t make,” Vafa explains.

Even though the transformers generated accurate directions and valid Othello moves in nearly every instance, the two metrics revealed that only one generated a coherent world model for Othello moves, and none performed well at forming coherent world models in the wayfinding example.

The researchers demonstrated the implications of this by adding detours to the map of New York City, which caused all the navigation models to fail.

“I was surprised by how quickly the performance deteriorated as soon as we added a detour. If we close just 1 percent of the possible streets, accuracy immediately plummets from nearly 100 percent to just 67 percent,” Vafa says.

When they recovered the city maps the models generated, they looked like an imagined New York City with hundreds of streets crisscrossing overlaid on top of the grid. The maps often contained random flyovers above other streets or multiple streets with impossible orientations.

These results show that transformers can perform surprisingly well at certain tasks without understanding the rules. If scientists want to build LLMs that can capture accurate world models, they need to take a different approach, the researchers say.

“Often, we see these models do impressive things and think they must have understood something about the world. I hope we can convince people that this is a question to think very carefully about, and we don’t have to rely on our own intuitions to answer it,” says Rambachan.

In the future, the researchers want to tackle a more diverse set of problems, such as those where some rules are only partially known. They also want to apply their evaluation metrics to real-world, scientific problems.

This work is funded, in part, by the Harvard Data Science Initiative, a National Science Foundation Graduate Research Fellowship, a Vannevar Bush Faculty Fellowship, a Simons Collaboration grant, and a grant from the MacArthur Foundation.

MIT EECS hosts workshop for Rising Stars in electrical engineering and computer science

Posted on November 5, 2024 by - News

Earlier this month, electrical engineering and computer science researchers from around the world came together at MIT for the twelfth annual Rising Stars Workshop. The event welcomed graduate students and postdocs of historically underrepresented genders who are interested in pursuing academic careers in the field.

When it first launched at MIT in 2012, the EECS Rising Stars Workshop aimed to help women scholars navigate the academic job search process and find the right match for their research. It has since been hosted at schools like the University of California at Berkeley, Carnegie Mellon University, and Stanford University. This year, the event returned to its beginnings, with Computer Science and Artificial Intelligence Laboratory (CSAIL) faculty helping lead the way.

“This program is an invaluable resource to the next generation of academic leaders in EECS,” says MIT Homer A. Burnell Career Development Assistant Professor, CSAIL principal investigator, and program co-chair Sara Beery. “It provides insight on how to navigate the faculty job market and the first years as a professor from diverse panels of successful faculty and research scientists from historically underrepresented genders. Perhaps even more valuable, however, is the network of mentors and peers they develop through the program, many of whom will remain close connections throughout their academic careers.”

The MIT-hosted program started off with workshop chair Dan Huttenlocher, dean of the MIT Schwarzman College of Computing, highlighting the new Building 45. “It’s a physical manifestation of the College’s three-fold mission: infusing the forefront of computing with fields across MIT, fortifying core computer science and artificial intelligence leadership, and advancing social, ethical, and policy dimensions of computing,” he said. Huttenlocher also emphasized the importance of broadening participation in computing to reflect diverse backgrounds and social needs.

Next, fellow workshop chair Asu Ozdaglar, deputy dean of the MIT Schwarzman College of Computing and head of the school’s Department of Electrical Engineering and Computer Science (EECS), discussed how MIT’s extensive efforts are helping expand how many women and minority-gender students major in EECS. “Over the last ten years, we have also been working deliberately on our hiring practices. As a result, each new crop of academic hires has been nearly 50% women for the last several years,” Ozdaglar noted. “Here at the beginning of your careers, the truth is that you’re all very powerful and can chart your own professional path.”

Each panel at the event focused on different stages of the faculty hiring process. The first discussion explored the application process, and included Yiyue Luo PhD ‘24, Assistant Professor at the University of Washington and former CSAIL affiliate. Together, they recommended researchers envision what they could contribute to a department before applying. “Consider whether you’d actually accept their offer before applying,” added Luo.

The workshop then shifted to lightning talks. Huddling into breakout rooms, the researchers introduced themselves in two-minute elevator pitches. For example, CSAIL PhD student Belinda Li presented her work on natural language processing, aiming to make language models more understandable, reliable, and user-friendly. Similarly, Princeton PhD student Sunnie S. Y. Kim talked about improving the transparency and fairness of AI models, while University of California San Diego PhD student Zih-Yun Chiu discussed how she’s using robots to help automate medical operations in dangerous or underserved areas.

At the poster sessions, Rising Stars like MIT EECS PhD student and CSAIL member Pratyusha Sharma discussed her latest research findings. She studies the role of language in intelligent behaviors across AI models, natural communication systems like those in whales, and robots. Recently, her team collaborated with Project CETI to use algorithms to decode the “sperm whale phonetic alphabet,” revealing sophisticated structures in their communicative clicks similar to human phonetics and the ways other animal species interact.

Yael Vinker, a postdoc in MIT professor and CSAIL principal investigator Antonio Torralba’s Vision Group, presented her work on the ways pretrained vision language models (VLMs) automatically produce visual elements. Making these designs clear and compelling is difficult for these models, but Vinker found that VLMs can, for example, combine aspects from different visuals into new images. She also observed that these models can generate clever typography (like a “Rising Stars” logo that uses a star symbol in place of the letter “A”).

Two subsequent panels explored how to navigate interviews and negotiate different offers. For the former, the speakers emphasized preparing a long-term research vision so prospective colleagues can better understand your goals. At the latter, panelists agreed that you shouldn’t be shy to negotiate for more students or summer salary, as examples. Beery chimed in with a valuable tip: “During the negotiation process, be sure to advocate for more compute for AI on behalf of your department.”

The day concluded with lab tours at Building 45. In Beery’s group, researchers showcased how they’re applying machine learning to environmental and biodiversity monitoring. By scaling up this data, her team can find the drivers of environmental changes in salmon habitats, for instance. Down the hall, researchers in MIT EECS associate professor and CSAIL principal investigator Pulkit Agrawal’s Improbable AI Lab demonstrated a walking humanoid robot. They noted that the machine is trained purely in simulation, and without seeing its surroundings, the robot still learns to walk by reading its joint angles.

On the second day, panelists explored topics like recruiting students and postdocs, making an impact, and funding. As one example of their valuable advice, speakers suggested that when you encounter unconscious biases within your research group, you can educate colleagues and foster more understanding about that gender dynamic. Between the panels, coffee breaks, and even a little mini golf, the Rising Stars took in valuable insights about academic careers in electrical engineering and computer science. Researchers walked away from the EECS Rising Stars Workshop feeling empowered by the connections they built and inspired by the conversations the event facilitated.

The workshop’s three other program co-chairs were X-Window Consortium Career Development Assistant Professor Tess Smidt and two CSAIL affiliates: KDD Career Development Professor in Communications and Technology Associate Professor Christina Delimitrou and TIBCO Career Development Associate Professor Stefanie Mueller. Anantha P. Chandrakasan, Dean of MIT’s School of Engineering and Vannevar Bush Professor of EECS, was a workshop chair for the event.

Quantum simulator could help uncover materials for high-performance electronics

Posted on October 31, 2024 by - News

Quantum computers hold the promise to emulate complex materials, helping researchers better understand the physical properties that arise from interacting atoms and electrons. This may one day lead to the discovery or design of better semiconductors, insulators, or superconductors that could be used to make ever faster, more powerful, and more energy-efficient electronics.

But some phenomena that occur in materials can be challenging to mimic using quantum computers, leaving gaps in the problems that scientists have explored with quantum hardware.

To fill one of these gaps, MIT researchers developed a technique to generate synthetic electromagnetic fields on superconducting quantum processors. The team demonstrated the technique on a processor comprising 16 qubits.

By dynamically controlling how the 16 qubits in their processor are coupled to one another, the researchers were able to emulate how electrons move between atoms in the presence of an electromagnetic field. Moreover, the synthetic electromagnetic field is broadly adjustable, enabling scientists to explore a range of material properties.

Emulating electromagnetic fields is crucial to fully explore the properties of materials. In the future, this technique could shed light on key features of electronic systems, such as conductivity, polarization, and magnetization.

“Quantum computers are powerful tools for studying the physics of materials and other quantum mechanical systems. Our work enables us to simulate much more of the rich physics that has captivated materials scientists,” says Ilan Rosen, an MIT postdoc and lead author of a paper on the quantum simulator.

The senior author is William D. Oliver, the Henry Ellis Warren professor of electrical engineering and computer science and of physics, director of the Center for Quantum Engineering, leader of the Engineering Quantum Systems group, and associate director of the Research Laboratory of Electronics. Oliver and Rosen are joined by others in the departments of Electrical Engineering and Computer Science and of Physics and at MIT Lincoln Laboratory. The research appears today in Nature Physics.

A quantum emulator

Companies like IBM and Google are striving to build large-scale digital quantum computers that hold the promise of outperforming their classical counterparts by running certain algorithms far more rapidly.

But that’s not all quantum computers can do. The dynamics of qubits and their couplings can also be carefully constructed to mimic the behavior of electrons as they move among atoms in solids.

“That leads to an obvious application, which is to use these superconducting quantum computers as emulators of materials,” says Jeffrey Grover, a research scientist at MIT and co-author on the paper.

Rather than trying to build large-scale digital quantum computers to solve extremely complex problems, researchers can use the qubits in smaller-scale quantum computers as analog devices to replicate a material system in a controlled environment.

“General-purpose digital quantum simulators hold tremendous promise, but they are still a long way off. Analog emulation is another approach that may yield useful results in the near-term, particularly for studying materials. It is a straightforward and powerful application of quantum hardware,” explains Rosen. “Using an analog quantum emulator, I can intentionally set a starting point and then watch what unfolds as a function of time.”

Despite their close similarity to materials, there are a few important ingredients in materials that can’t be easily reflected on quantum computing hardware. One such ingredient is a magnetic field.

In materials, electrons “live” in atomic orbitals. When two atoms are close to one another, their orbitals overlap and electrons can “hop” from one atom to another. In the presence of a magnetic field, that hopping behavior becomes more complex.

On a superconducting quantum computer, microwave photons hopping between qubits are used to mimic electrons hopping between atoms. But, because photons are not charged particles like electrons, the photons’ hopping behavior would remain the same in a physical magnetic field.

Since they can’t just turn on a magnetic field in their simulator, the MIT team employed a few tricks to synthesize the effects of one instead.

Tuning up the processor

The researchers adjusted how adjacent qubits in the processor were coupled to each other to create the same complex hopping behavior that electromagnetic fields cause in electrons.

To do that, they slightly changed the energy of each qubit by applying different microwave signals. Usually, researchers will set qubits to the same energy so that photons can hop from one to another. But for this technique, they dynamically varied the energy of each qubit to change how they communicate with each other.

By precisely modulating these energy levels, the researchers enabled photons to hop between qubits in the same complex manner that electrons hop between atoms in a magnetic field.

Plus, because they can finely tune the microwave signals, they can emulate a range of electromagnetic fields with different strengths and distributions.

The researchers undertook several rounds of experiments to determine what energy to set for each qubit, how strongly to modulate them, and the microwave frequency to use.

“The most challenging part was finding modulation settings for each qubit so that all 16 qubits work at once,” Rosen says.

Once they arrived at the right settings, they confirmed that the dynamics of the photons uphold several equations that form the foundation of electromagnetism. They also demonstrated the “Hall effect,” a conduction phenomenon that exists in the presence of an electromagnetic field.

These results show that their synthetic electromagnetic field behaves like the real thing.

Moving forward, they could use this technique to precisely study complex phenomena in condensed matter physics, such as phase transitions that occur when a material changes from a conductor to an insulator.

“A nice feature of our emulator is that we need only change the modulation amplitude or frequency to mimic a different material system. In this way, we can scan over many materials properties or model parameters without having to physically fabricate a new device each time.” says Oliver.

While this work was an initial demonstration of a synthetic electromagnetic field, it opens the door to many potential discoveries, Rosen says.

“The beauty of quantum computers is that we can look at exactly what is happening at every moment in time on every qubit, so we have all this information at our disposal. We are in a very exciting place for the future,” he adds.

This work is supported, in part, by the U.S. Department of Energy, the U.S. Defense Advanced Research Projects Agency (DARPA), the U.S. Army Research Office, the Oak Ridge Institute for Science and Education, the Office of the Director of National Intelligence, NASA, and the National Science Foundation. 

Interactive mouthpiece opens new opportunities for health data, assistive technology, and hands-free interactions

Posted on October 30, 2024 by - News

When you think about hands-free devices, you might picture Alexa and other voice-activated in-home assistants, Bluetooth earpieces, or asking Siri to make a phone call in your car. You might not imagine using your mouth to communicate with other devices like a computer or a phone remotely. 

Thinking outside the box, MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) and Aarhus University researchers have now engineered “MouthIO,” a dental brace that can be fabricated with sensors and feedback components to capture in-mouth interactions and data. This interactive wearable could eventually assist dentists and other doctors with collecting health data and help motor-impaired individuals interact with a phone, computer, or fitness tracker using their mouths.

Resembling an electronic retainer, MouthIO is a see-through brace that fits the specifications of your upper or lower set of teeth from a scan. The researchers created a plugin for the modeling software Blender to help users tailor the device to fit a dental scan, where you can then 3D print your design in dental resin. This computer-aided design tool allows users to digitally customize a panel (called PCB housing) on the side to integrate electronic components like batteries, sensors (including detectors for temperature and acceleration, as well as tongue-touch sensors), and actuators (like vibration motors and LEDs for feedback). You can also place small electronics outside of the PCB housing on individual teeth.

The active mouth

“The mouth is a really interesting place for an interactive wearable and can open up many opportunities, but has remained largely unexplored due to its complexity,” says senior author Michael Wessely, a former CSAIL postdoc and senior author on a paper about MouthIO who is now an assistant professor at Aarhus University. “This compact, humid environment has elaborate geometries, making it hard to build a wearable interface to place inside. With MouthIO, though, we’ve developed a new kind of device that’s comfortable, safe, and almost invisible to others. Dentists and other doctors are eager about MouthIO for its potential to provide new health insights, tracking things like teeth grinding and potentially bacteria in your saliva.”

The excitement for MouthIO’s potential in health monitoring stems from initial experiments. The team found that their device could track bruxism (the habit of grinding teeth) by embedding an accelerometer within the brace to track jaw movements. When attached to the lower set of teeth, MouthIO detected when users grind and bite, with the data charted to show how often users did each.

Wessely and his colleagues’ customizable brace could one day help users with motor impairments, too. The team connected small touchpads to MouthIO, helping detect when a user’s tongue taps their teeth. These interactions could be sent via Bluetooth to scroll across a webpage, for example, allowing the tongue to act as a “third hand” to open up a new avenue for hands-free interaction.

“MouthIO is a great example how miniature electronics now allow us to integrate sensing into a broad range of everyday interactions,” says study co-author Stefanie Mueller, the TIBCO Career Development Associate Professor in the MIT departments of Electrical Engineering and Computer Science and Mechanical Engineering and leader of the HCI Engineering Group at CSAIL. “I’m especially excited about the potential to help improve accessibility and track potential health issues among users.”

Molding and making MouthIO

To get a 3D model of your teeth, you can first create a physical impression and fill it with plaster. You can then scan your mold with a mobile app like Polycam and upload that to Blender. Using the researchers’ plugin within this program, you can clean up your dental scan to outline a precise brace design. Finally, you 3D print your digital creation in clear dental resin, where the electronic components can then be soldered on. Users can create a standard brace that covers their teeth, or opt for an “open-bite” design within their Blender plugin. The latter fits more like open-finger gloves, exposing the tips of your teeth, which helps users avoid lisping and talk naturally.

This “do it yourself” method costs roughly $15 to produce and takes two hours to be 3D-printed. MouthIO can also be fabricated with a more expensive, professional-level teeth scanner similar to what dentists and orthodontists use, which is faster and less labor-intensive.

Compared to its closed counterpart, which fully covers your teeth, the researchers view the open-bite design as a more comfortable option. The team preferred to use it for beverage monitoring experiments, where they fabricated a brace capable of alerting users when a drink was too hot. This iteration of MouthIO had a temperature sensor and a monitor embedded within the PCB housing that vibrated when a drink exceeded 65 degrees Celsius (or 149 degrees Fahrenheit). This could help individuals with mouth numbness better understand what they’re consuming.

In a user study, participants also preferred the open-bite version of MouthIO. “We found that our device could be suitable for everyday use in the future,” says study lead author and Aarhus University PhD student Yijing Jiang. “Since the tongue can touch the front teeth in our open-bite design, users don’t have a lisp. This made users feel more comfortable wearing the device during extended periods with breaks, similar to how people use retainers.”

The team’s initial findings indicate that MouthIO is a cost-effective, accessible, and customizable interface, and the team is working on a more long-term study to evaluate its viability further. They’re looking to improve its design, including experimenting with more flexible materials, and placing it in other parts of the mouth, like the cheek and the palate. Among these ideas, the researchers have already prototyped two new designs for MouthIO: a single-sided brace for even higher comfort when wearing MouthIO while also being fully invisible to others, and another fully capable of wireless charging and communication.

Jiang, Mueller, and Wessely’s co-authors include PhD student Julia Kleinau, master’s student Till Max Eckroth, and associate professor Eve Hoggan, all of Aarhus University. Their work was supported by a Novo Nordisk Foundation grant and was presented at ACM’s Symposium on User Interface Software and Technology.

A faster, better way to train general-purpose robots

Posted on October 30, 2024 by - News

In the classic cartoon “The Jetsons,” Rosie the robotic maid seamlessly switches from vacuuming the house to cooking dinner to taking out the trash. But in real life, training a general-purpose robot remains a major challenge.

Typically, engineers collect data that are specific to a certain robot and task, which they use to train the robot in a controlled environment. However, gathering these data is costly and time-consuming, and the robot will likely struggle to adapt to environments or tasks it hasn’t seen before.

To train better general-purpose robots, MIT researchers developed a versatile technique that combines a huge amount of heterogeneous data from many of sources into one system that can teach any robot a wide range of tasks.

Their method involves aligning data from varied domains, like simulations and real robots, and multiple modalities, including vision sensors and robotic arm position encoders, into a shared “language” that a generative AI model can process.

By combining such an enormous amount of data, this approach can be used to train a robot to perform a variety of tasks without the need to start training it from scratch each time.

This method could be faster and less expensive than traditional techniques because it requires far fewer task-specific data. In addition, it outperformed training from scratch by more than 20 percent in simulation and real-world experiments.

“In robotics, people often claim that we don’t have enough training data. But in my view, another big problem is that the data come from so many different domains, modalities, and robot hardware. Our work shows how you’d be able to train a robot with all of them put together,” says Lirui Wang, an electrical engineering and computer science (EECS) graduate student and lead author of a paper on this technique.

Wang’s co-authors include fellow EECS graduate student Jialiang Zhao; Xinlei Chen, a research scientist at Meta; and senior author Kaiming He, an associate professor in EECS and a member of the Computer Science and Artificial Intelligence Laboratory (CSAIL). The research will be presented at the Conference on Neural Information Processing Systems.

Inspired by LLMs

A robotic “policy” takes in sensor observations, like camera images or proprioceptive measurements that track the speed and position a robotic arm, and then tells a robot how and where to move.

Policies are typically trained using imitation learning, meaning a human demonstrates actions or teleoperates a robot to generate data, which are fed into an AI model that learns the policy. Because this method uses a small amount of task-specific data, robots often fail when their environment or task changes.

To develop a better approach, Wang and his collaborators drew inspiration from large language models like GPT-4.

These models are pretrained using an enormous amount of diverse language data and then fine-tuned by feeding them a small amount of task-specific data. Pretraining on so much data helps the models adapt to perform well on a variety of tasks.

“In the language domain, the data are all just sentences. In robotics, given all the heterogeneity in the data, if you want to pretrain in a similar manner, we need a different architecture,” he says.

Robotic data take many forms, from camera images to language instructions to depth maps. At the same time, each robot is mechanically unique, with a different number and orientation of arms, grippers, and sensors. Plus, the environments where data are collected vary widely.

The MIT researchers developed a new architecture called Heterogeneous Pretrained Transformers (HPT) that unifies data from these varied modalities and domains.

They put a machine-learning model known as a transformer into the middle of their architecture, which processes vision and proprioception inputs. A transformer is the same type of model that forms the backbone of large language models.

The researchers align data from vision and proprioception into the same type of input, called a token, which the transformer can process. Each input is represented with the same fixed number of tokens.

Then the transformer maps all inputs into one shared space, growing into a huge, pretrained model as it processes and learns from more data. The larger the transformer becomes, the better it will perform.

A user only needs to feed HPT a small amount of data on their robot’s design, setup, and the task they want it to perform. Then HPT transfers the knowledge the transformer grained during pretraining to learn the new task.

Enabling dexterous motions

One of the biggest challenges of developing HPT was building the massive dataset to pretrain the transformer, which included 52 datasets with more than 200,000 robot trajectories in four categories, including human demo videos and simulation.

The researchers also needed to develop an efficient way to turn raw proprioception signals from an array of sensors into data the transformer could handle.

“Proprioception is key to enable a lot of dexterous motions. Because the number of tokens is in our architecture always the same, we place the same importance on proprioception and vision,” Wang explains.

When they tested HPT, it improved robot performance by more than 20 percent on simulation and real-world tasks, compared with training from scratch each time. Even when the task was very different from the pretraining data, HPT still improved performance.

“This paper provides a novel approach to training a single policy across multiple robot embodiments. This enables training across diverse datasets, enabling robot learning methods to significantly scale up the size of datasets that they can train on. It also allows the model to quickly adapt to new robot embodiments, which is important as new robot designs are continuously being produced,” says David Held, associate professor at the Carnegie Mellon University Robotics Institute, who was not involved with this work.

In the future, the researchers want to study how data diversity could boost the performance of HPT. They also want to enhance HPT so it can process unlabeled data like GPT-4 and other large language models.

“Our dream is to have a universal robot brain that you could download and use for your robot without any training at all. While we are just in the early stages, we are going to keep pushing hard and hope scaling leads to a breakthrough in robotic policies, like it did with large language models,” he says.

This work was funded, in part, by the Amazon Greater Boston Tech Initiative and the Toyota Research Institute.

Making it easier to verify an AI model’s responses

Posted on October 25, 2024 by - News

Despite their impressive capabilities, large language models are far from perfect. These artificial intelligence models sometimes “hallucinate” by generating incorrect or unsupported information in response to a query.

Due to this hallucination problem, an LLM’s responses are often verified by human fact-checkers, especially if a model is deployed in a high-stakes setting like health care or finance. However, validation processes typically require people to read through long documents cited by the model, a task so onerous and error-prone it may prevent some users from deploying generative AI models in the first place.

To help human validators, MIT researchers created a user-friendly system that enables people to verify an LLM’s responses much more quickly. With this tool, called SymGen, an LLM generates responses with citations that point directly to the place in a source document, such as a given cell in a database.

Users hover over highlighted portions of its text response to see data the model used to generate that specific word or phrase. At the same time, the unhighlighted portions show users which phrases need additional attention to check and verify.

“We give people the ability to selectively focus on parts of the text they need to be more worried about. In the end, SymGen can give people higher confidence in a model’s responses because they can easily take a closer look to ensure that the information is verified,” says Shannon Shen, an electrical engineering and computer science graduate student and co-lead author of a paper on SymGen.

Through a user study, Shen and his collaborators found that SymGen sped up verification time by about 20 percent, compared to manual procedures. By making it faster and easier for humans to validate model outputs, SymGen could help people identify errors in LLMs deployed in a variety of real-world situations, from generating clinical notes to summarizing financial market reports.

Shen is joined on the paper by co-lead author and fellow EECS graduate student Lucas Torroba Hennigen; EECS graduate student Aniruddha “Ani” Nrusimha; Bernhard Gapp, president of the Good Data Initiative; and senior authors David Sontag, a professor of EECS, a member of the MIT Jameel Clinic, and the leader of the Clinical Machine Learning Group of the Computer Science and Artificial Intelligence Laboratory (CSAIL); and Yoon Kim, an assistant professor of EECS and a member of CSAIL. The research was recently presented at the Conference on Language Modeling.

Symbolic references

To aid in validation, many LLMs are designed to generate citations, which point to external documents, along with their language-based responses so users can check them. However, these verification systems are usually designed as an afterthought, without considering the effort it takes for people to sift through numerous citations, Shen says.

“Generative AI is intended to reduce the user’s time to complete a task. If you need to spend hours reading through all these documents to verify the model is saying something reasonable, then it’s less helpful to have the generations in practice,” Shen says.

The researchers approached the validation problem from the perspective of the humans who will do the work.

A SymGen user first provides the LLM with data it can reference in its response, such as a table that contains statistics from a basketball game. Then, rather than immediately asking the model to complete a task, like generating a game summary from those data, the researchers perform an intermediate step. They prompt the model to generate its response in a symbolic form.

With this prompt, every time the model wants to cite words in its response, it must write the specific cell from the data table that contains the information it is referencing. For instance, if the model wants to cite the phrase “Portland Trailblazers” in its response, it would replace that text with the cell name in the data table that contains those words.

“Because we have this intermediate step that has the text in a symbolic format, we are able to have really fine-grained references. We can say, for every single span of text in the output, this is exactly where in the data it corresponds to,” Torroba Hennigen says.

SymGen then resolves each reference using a rule-based tool that copies the corresponding text from the data table into the model’s response.

“This way, we know it is a verbatim copy, so we know there will not be any errors in the part of the text that corresponds to the actual data variable,” Shen adds.

Streamlining validation

The model can create symbolic responses because of how it is trained. Large language models are fed reams of data from the internet, and some data are recorded in “placeholder format” where codes replace actual values.

When SymGen prompts the model to generate a symbolic response, it uses a similar structure.

“We design the prompt in a specific way to draw on the LLM’s capabilities,” Shen adds.

During a user study, the majority of participants said SymGen made it easier to verify LLM-generated text. They could validate the model’s responses about 20 percent faster than if they used standard methods.

However, SymGen is limited by the quality of the source data. The LLM could cite an incorrect variable, and a human verifier may be none-the-wiser.

In addition, the user must have source data in a structured format, like a table, to feed into SymGen. Right now, the system only works with tabular data.

Moving forward, the researchers are enhancing SymGen so it can handle arbitrary text and other forms of data. With that capability, it could help validate portions of AI-generated legal document summaries, for instance. They also plan to test SymGen with physicians to study how it could identify errors in AI-generated clinical summaries.

This work is funded, in part, by Liberty Mutual and the MIT Quest for Intelligence Initiative.

Using spatial learning to transform math and science education

Posted on October 18, 2024 by - News

Legend has it that Isaac Newton was sitting under a tree when an apple fell on his head, sparking a bout of scientific thinking that led to the theory of gravity. It’s one of the most famous stories in science, perhaps because it shows the power of simple human experiences to revolutionize our understanding of the world around us.

About five years ago, Anurupa Ganguly ’07, MNG ’09 noticed kids don’t learn that way in schools.

“Students should learn how to use language, notation, and eventually shorthand representation of thoughts from deeply human experiences,” Ganguly says.

That’s the idea behind PrismsVR. The company offers virtual reality experiences for students, using physical learning to teach core concepts in math and science.

The platform can radically change the dynamics of the classroom, encouraging self-paced, student-led learning, where the teacher is focused on asking the right questions and sparking curiosity.

Instead of learning biology with a pen and paper, students become biomedical researchers designing a tissue regeneration therapy. Instead of learning trigonometry in a textbook, students become rural architects designing a new school building.

“We’re building a whole new learning platform, methodology, and tech infrastructure that allows students to experience problems in the first person, not through abstractions or 2D screens, and then go from that experience to ascribe meaning, language, and build up to equations, procedures, and other nomenclature,” Ganguly explains.

A 3D line chart has lines going up and down in green and red.
Students can rotate their bodies to minimize the distance between a trend line and the data points, to find the line of best fit and formalize the concept of residuals.

Image: Courtesy of PrismsVR

Today PrismsVR has been used by about 300,000 students across 35 states. The company’s approach was shown to boost algebra test scores by 11 percent in one study, with larger, multistate studies currently underway through funding from the Gates Foundation.

“Education has been in desperate need of real reform for many years,” Ganguly says. “But what’s happened is we’ve just been digitizing old, antiquated teaching methods instead. We would take a lecture and make it a video, or take a worksheet and make it a web app. I think districts see us taking a more aspirational approach, with multimodal interaction and concepts at the center of learning design, and are collaborating with us to scale that instead. We want to get this to every single public school student across the U.S., and then we’re going into community colleges, higher ed, and international.”

A new paradigm for learning

Ganguly was an undergraduate and master’s student in MIT’s Department of Electrical Engineering and Computer Science. When she began as an undergrad in 2003, she estimates that women made up about 30 percent of her class in the department, but as she advanced in her studies, that number seemed to dwindle.

“It was a disappearing act for some students, and I became inspired to understand what’s happening at the K-12 levels that set some students up for success and led to fragile foundations for others,” Ganguly recalls.

As she neared the end of her graduate program in 2009, Ganguly planned to move to California to take an engineering job. But as she was walking through MIT’s Infinite Corridor one day, a sign caught her eye. It was for Teach for America, which had collaborated with MIT to recruit students into the field of teaching, particularly for high need and high poverty students.

“I was inspired by that idea that I could use my education, engineering background, and disciplined systems thinking to think through systemic change in the public sector,” says Ganguly, who became a high school physics and algebra teacher in the Boston Public Schools.

Ganguly soon left the classroom and became director of math for the district, where she oversaw curriculum and teacher upskilling. From there, Ganguly went to New York City Public Schools, where she also supported curriculum development, trying to relate abstract math concepts to students’ experiences in the real world.

“As I began to travel from school to school, working with millions of kids, I became convinced that we don’t have the tools to solve the problem I thought about at MIT — of truly leveling the playing field and building enduring identities in the mathematical sciences,” Ganguly says.

The problem as Ganguly sees it is that students’ world is 3D, complex, and multimodal. Yet most lessons are confined to paper or tablets. For other things in life, students learn through their complex experiences: through their senses, movement, and emotions. Why should math and science be any different? In 2018, the Oculus Quest VR headset was released, and Ganguly thought she had found a more effective learning medium to scale how we learn.

But starting an education company based on virtual reality at the time was audacious. The 128-gigabyte Quest was priced at $500, and there were no standards-based VR curricula or standalone VR headsets in U.S. K-12 schools.

“Investors weren’t going to touch this,” Ganguly jokes.

Luckily, Ganguly received a small amount of funding from the National Science Foundation to build her first prototype. Ganguly started with Algebra 1; performance in this class is one of the top predictors of lifetime wages but has shown a stubbornly persistent achievement gap.

Her first module, which she built during the pandemic, places students in a food hall when a sudden announcement from the mayor rings out. There’s an alarming growth of an unknown virus in the area. The students get the power to travel back in time to see how the virus is spreading, from one person’s sneeze to many people’s behaviors in a demonstration of multiplicative growth.

The people turn to dots in a simulation as the journey moves to interactive, tactile data visualization, and the students are charged with figuring out how many weeks until the hospitals run out of capacity. Once the learning design for VR was established, Ganguly continued to build experiences across the curriculum in geometry, algebra II and III, biology, chemistry, and middle school subjects. Today Prisms covers all math and science subjects in grades seven to eleven, and the company is currently building out calculus, data science, and statistics for upper and postsecondary school. By the fall of 2025, Prisms will have evergreen content up to grade level 14.

Following the experiences, students gather in small groups to reflect on the lessons and write summaries. As students go through their virtual experiences, teachers have a web dashboard to monitor each child’s progress to support and intervene where needed.

“With our solution, the role of the teacher is to be Socrates and to ask high-quality questions, not deliver knowledge” Ganguly says.

As a solo founder, Ganguly says support from MIT’s Venture Mentoring Service, which offers members of the MIT community startup guidance in the form of “board meetings” led by successful entrepreneurs, was crucial.

“The MIT founder community is different,” Ganguly says. “We’re often technical founders, building for ourselves, and we build our company’s first product. Moving from product to your go-to-market strategy and hiring is a unique journey for product-minded founders.”

“We intend to become the next textbook,” Anurupa Ganguly, left, says. “The next textbooks will be spatial and experiential.” Image courtesy PrismsVR

From textbooks to experiences

A few years ago, Ganguly’s team was leading a classroom coaching session in a Virginia school district when a teacher told her about a student named Silas.

“The teacher was saying, ‘Silas never does anything, he just sits in the back of class,’” Ganguly recalls. “I’ve seen this like clockwork, so we just said, ‘Let’s give Silas a fresh shot and see what we can do.’ Lo and behold, Silas was the first one to finish the module and write a full synthesis report. The teacher told me that was the first time Silas has turned in an assignment with everything filled in.”

Ganguly says it’s one of thousands of anecdotes she has.

“A lot of students feel shut out of the modern math classroom because of our stubborn approach of drill and kill,” Ganguly says. “Students want to learn through great stories. They want to help people. They want to be empathetic. They want their math education to matter.”

Ganguly sees PrismsVR as a fundamentally new way for students to learn no matter where they are.

“We intend to become the next textbook,” Ganguly says. “The next textbooks will be spatial and experiential.”