Page 29 – MIT EECS

MIT researchers introduce generative AI for databases

Posted on July 11, 2024 by Jane Halpern - News

A new tool makes it easier for database users to perform complicated statistical analyses of tabular data without the need to know what is going on behind the scenes.

GenSQL, a generative AI system for databases, could help users make predictions, detect anomalies, guess missing values, fix errors, or generate synthetic data with just a few keystrokes.

For instance, if the system were used to analyze medical data from a patient who has always had high blood pressure, it could catch a blood pressure reading that is low for that particular patient but would otherwise be in the normal range.

GenSQL automatically integrates a tabular dataset and a generative probabilistic AI model, which can account for uncertainty and adjust their decision-making based on new data.

Moreover, GenSQL can be used to produce and analyze synthetic data that mimic the real data in a database. This could be especially useful in situations where sensitive data cannot be shared, such as patient health records, or when real data are sparse.

This new tool is built on top of SQL, a programming language for database creation and manipulation that was introduced in the late 1970s and is used by millions of developers worldwide.

“Historically, SQL taught the business world what a computer could do. They didn’t have to write custom programs, they just had to ask questions of a database in high-level language. We think that, when we move from just querying data to asking questions of models and data, we are going to need an analogous language that teaches people the coherent questions you can ask a computer that has a probabilistic model of the data,” says Vikash Mansinghka ’05, MEng ’09, PhD ’09, senior author of a paper introducing GenSQL and a principal research scientist and leader of the Probabilistic Computing Project in the MIT Department of Brain and Cognitive Sciences.

When the researchers compared GenSQL to popular, AI-based approaches for data analysis, they found that it was not only faster but also produced more accurate results. Importantly, the probabilistic models used by GenSQL are explainable, so users can read and edit them.

“Looking at the data and trying to find some meaningful patterns by just using some simple statistical rules might miss important interactions. You really want to capture the correlations and the dependencies of the variables, which can be quite complicated, in a model. With GenSQL, we want to enable a large set of users to query their data and their model without having to know all the details,” adds lead author Mathieu Huot, a research scientist in the Department of Brain and Cognitive Sciences and member of the Probabilistic Computing Project.

They are joined on the paper by Matin Ghavami and Alexander Lew, MIT graduate students; Cameron Freer, a research scientist; Ulrich Schaechtle and Zane Shelby of Digital Garage; Martin Rinard, an MIT professor in the Department of Electrical Engineering and Computer Science and member of the Computer Science and Artificial Intelligence Laboratory (CSAIL); and Feras Saad ’15, MEng ’16, PhD ’22, an assistant professor at Carnegie Mellon University. The research was recently presented at the ACM Conference on Programming Language Design and Implementation.

Combining models and databases

SQL, which stands for structured query language, is a programming language for storing and manipulating information in a database. In SQL, people can ask questions about data using keywords, such as by summing, filtering, or grouping database records.

However, querying a model can provide deeper insights, since models can capture what data imply for an individual. For instance, a female developer who wonders if she is underpaid is likely more interested in what salary data mean for her individually than in trends from database records.

The researchers noticed that SQL didn’t provide an effective way to incorporate probabilistic AI models, but at the same time, approaches that use probabilistic models to make inferences didn’t support complex database queries.

They built GenSQL to fill this gap, enabling someone to query both a dataset and a probabilistic model using a straightforward yet powerful formal programming language.

A GenSQL user uploads their data and probabilistic model, which the system automatically integrates. Then, she can run queries on data that also get input from the probabilistic model running behind the scenes. This not only enables more complex queries but can also provide more accurate answers.

For instance, a query in GenSQL might be something like, “How likely is it that a developer from Seattle knows the programming language Rust?” Just looking at a correlation between columns in a database might miss subtle dependencies. Incorporating a probabilistic model can capture more complex interactions.

Plus, the probabilistic models GenSQL utilizes are auditable, so people can see which data the model uses for decision-making. In addition, these models provide measures of calibrated uncertainty along with each answer.

For instance, with this calibrated uncertainty, if one queries the model for predicted outcomes of different cancer treatments for a patient from a minority group that is underrepresented in the dataset, GenSQL would tell the user that it is uncertain, and how uncertain it is, rather than overconfidently advocating for the wrong treatment.

Faster and more accurate results

To evaluate GenSQL, the researchers compared their system to popular baseline methods that use neural networks. GenSQL was between 1.7 and 6.8 times faster than these approaches, executing most queries in a few milliseconds while providing more accurate results.

They also applied GenSQL in two case studies: one in which the system identified mislabeled clinical trial data and the other in which it generated accurate synthetic data that captured complex relationships in genomics.

Next, the researchers want to apply GenSQL more broadly to conduct largescale modeling of human populations. With GenSQL, they can generate synthetic data to draw inferences about things like health and salary while controlling what information is used in the analysis.

They also want to make GenSQL easier to use and more powerful by adding new optimizations and automation to the system. In the long run, the researchers want to enable users to make natural language queries in GenSQL. Their goal is to eventually develop a ChatGPT-like AI expert one could talk to about any database, which grounds its answers using GenSQL queries.

This research is funded, in part, by the Defense Advanced Research Projects Agency (DARPA), Google, and the Siegel Family Foundation.

Best picnic spot near Cambridge–Killian Court, MIT!

Posted on July 9, 2024 by Jane Halpern - News

On a day that could only be described as “an 11 on a scale of 10”, Killian Court was a perfect spot for a picnic lunch! The EECS department head along with EECS faculty, staff, and current graduate students welcomed 16 undergraduate interns visiting MIT as part of the MIT Summer Research Program (MSRP) on July 3rd.

MSRP is a large, and growing, summer research program spanning 9 weeks (June 2-August 11) where 97 undergraduate interns are matched with a graduate student mentor and hosted by MIT faculty in various laboratories across campus. The program is intense with interns expected to work in the lab 40 hours per week, attend departmental information sessions, participate in professional development activities, and attend research seminars. Interns work hard in the lab on their own research projects with communication deliverables including a research proposal, progress reports, and a ﬁnal research poster highlighting their research ﬁndings; in their free time, interns even work on crafting various essays that will be part of their application to graduate school due later in the fall.

Department Head Asu Ozdaglar relaxes with MSRP participants.

EECS faculty are hosting 10 MSRP interns this summer, and so far, the department has offered:

an information session highlighting the doctoral graduate application process with the chair of graduate admissions
a conversation with the EECS department head about the research directions of EECS describing where the department is heading
an upcoming in-person laboratory tour of the Quantum Systems Laboratory
and, of course, a relaxing conversation during the picnic lunch. (Everyone must take a break for lunch and fortunately the MSRP interns said yes to our invitation!)

When posed the question “now that you are living and working at MIT, what surprised you about this place, or what is different than what you anticipated about MIT?” The replies from the interns included: “Everyone is so nice…from faculty all the way through the janitorial staff…everyone is willing to lend a hand in even the smallest circumstances.” “The transparency at MIT is great… mentors are very willing to talk about how the journey through research is not a straight line of success and they are willing to share their own experiences.” “I expected a very serious and stuffy environment, but everyone is very welcoming and collaborative.” Other comments included: “It’s okay to ask questions and ask for help.” “It is amazing that everyone in the lab knows your name.”

Of course, MIT has its downsides with remarks like “Cement buildings have no romance.” “There are mice everywhere!” “There are so few restaurants by my dorm (Burton-Connor), and I have to walk all the way to Kendall to get food.”

With our current PhD student hosts, the conversation included what it is like being a graduate student. Lots of great advice was shared, showing that great determination is needed and resilience too, as the journey is long (like a marathon). A grad student remarked “you get a PhD when you know something that no one else in the world knows, and you are the only one that has ﬁgured out the solution.” The current graduate students urged balance too; make sure to take care of yourself and to have fun as there are lots of great things to do in New England.

In sum–a short break on a perfect summer day at MIT enjoying a picnic on Killian Court with colleagues and friends!

Now, back to the lab; there is more research to do!

From group stretches to “Hitting Roman,” MIT Motorsports traditions live on

Posted on July 2, 2024 by Jane Halpern - News

While siblings Kevin Chan ’17 and rising senior Monica Chan may be seven years apart in age, as Monica Chan puts it, “we’re eight grades apart, so, like, eight life-years apart.”

Despite this age gap — Kevin left for college when Monica was in fifth grade — the siblings share remarkably similar experiences and interests. Both led subteams on the MIT Motorsports team, albeit eight years apart. Kevin was the electrical systems lead from 2015 to 2017, and Monica is the current software lead.

Founded in 2001 by Rich James ’04, SM ’06 and Nick Gidwani ’04, and supported by the Edgerton Center, MIT Motorsports designs and builds a high-caliber Formula SAE car to race in yearly competitions. Over the past 23 years, MIT Motorsports has built 19 cars, won 10 trophies, and has had hundreds of team members. Alumni are die-hard fans and established an endowed fund for their 20th anniversary to ensure the team’s longevity. In 2017, Kevin’s team won Second Place Overall at the Formula SAE Electric competition in Lincoln, Nebraska.

Kevin was one of two electrical engineering students on the team, and today Monica oversees a subteam of 10 students. The subteam expansion has facilitated the development of a custom telemetry system. “You can view live data coming off of the car that’s transmitted through radio, and we have a custom dashboard that we created with a custom PCB that transmits all that data now,” Monica says.

“It’s so funny to hear Monica talking about this, because when I was on the team, our UI [user interface] for the driver and everything was so simple. It was just a little, single-line display that showed the max cell temperature and minimum cell voltage,” Kevin chuckles. “And then we literally had a sticky note on the dashboard that was like, do not go above this temperature. Do not go below this voltage.”

While at MIT, Kevin kept up with his sister weekly, updating her on everything happening at Formula Society of Automotive Engineers (FSAE). “A big piece of advice Kevin gave me when I was a junior in high school was that you’re never too young to do something amazing,” Monica says. “He told me back then that ‘you’re not going to be much smarter two years from now than you are now.’ That piece of advice helped me get through high school and pushed me to do my best to do the hard and difficult things because indeed, it’s more about the personal qualities you have that push you to do the hard projects. Knowledge can always be acquired, but the drive is the harder part.”

A diverse team of college students gather around a red and white racecar with a high rear spoiler. In the foreground, a silver cup sits on the asphalt. — The MIT Motorsports team is all smiles with their Second Place Overall Spirit of Excellence Award silver trophy at the Formula SAE Electric competition in Lincoln, Nebraska. Photo courtesy of MIT Motorsports.

Traditions are part of the fabric of the team culture. Their team stretch at the end of every meeting is an enduring tradition. “Everyone just extends their arms out while standing up and then does a squat. Then, they clap. This is just a thing that has been done on the team since before I was on the team. They said that the origin of it was the stretch that Japanese autoworkers do at the beginning of the day to stretch out their jumpsuits in the factory and make the knees a little bit more flexible. And it’s just fascinating, because this stretch is now almost 20 years old on the team,” Kevin says.

“Hitting Roman,” the day the car first rolls, is an important milestone. “When I was on the team, we were convinced that saying that the car was going to run was bad luck,” Kevin says. “We were trying to come up with a new term to replace the term ‘running car’ because we thought that saying the words ‘running car’ would make it so that the car never ran. So instead of calling it a running car, we called it ‘Roman Chariot.’” The name stuck, and Monica’s team hit Roman in April.

For Kevin, the spirit of Motorsports remains ever-present, as he shares his home with four Motorsports alums and collaborates with three Motorsports alums at Tesla, where he serves as a staff energy systems design/architecture engineer.

“FSAE and the Edgerton Center played a huge role in jump starting my career and my internships. I think there’s not many places where you can get both the breadth and the depth of the design process,” Kevin says.

For Monica, “Race car puts many things in perspective where you implement a lot of the things that you learn in class into a physical project. Sometimes I learn things through race car before I learn them in class. And then when I go back to class, it gives me a better physical intuition for how something works because I have experience implementing it.”

The team recently returned from the Formula Hybrid competition in Loudon, New Hampshire, where they finished first in design, first in scrutineering [mandatory technical, safety, and administrative checks], second in acceleration, third in the racing challenge, fourth in project management, and fifth overall. Edgerton Center Technical Instructor Pat McAtamney reports, “I’ve never seen a team complete a brakes test in one try.”

An assisted step forward: Max Burns talks about the process, and teamwork, of invention

Posted on July 1, 2024 by Jane Halpern - News

When people hear the term “inventor”, they often picture someone working alone: Ben Franklin and his kite; Edison and his lightbulb. But Max Burns would like you to picture something else: a team.

The SuperUROP participant from Logan, Ohio just walked with his degree in Mechanical Engineering this spring. Along the way, Burns’s work has been recognized by multiple entities: as an undergraduate, he was a recipient of the Arthur J. Samberg scholarship, and of the Prince Innovation Prize from the Mechanical Engineering Department for his SuperUROP project, an instrumented walking cane. For his coming stint in graduate school at Stanford University, Burns has already received the NSF Graduate Research Fellowship, the Stanford Graduate Fellowship, and the Stanford EDGE (Enhancing Diversity in Graduate Education) Fellowship. All this support is fitting, because Max Burns is all about giving people the support they need. We sat down with him to learn more.

For your SuperUROP project, you built a walking cane with sensing instruments that could gather data about a patient’s balance as they used it. Where did you begin with the development of this project—what inspired it?

I began this project the year before joining SuperUROP, when I was looking for a UROP experience where I could explore assistive technology research. I joined the MIT Newman Laboratory for Human Rehabilitation, where I’ve been advised by graduate student Kaymie Shiozawa and Professor Neville Hogan. I was excited to work with the instrumented cane, because my grandmother relies on a walking cane and struggles with balance issues and falls. After I learned more about balance, I better understood how dangerous falling is for older adults, and the way mobility limitations take away independence and limit millions of people’s personal freedom. From what I learned during this project, I’ve decided to focus on lower limb mobility assistance in my graduate work.

In terms of the project’s development, the cane was actually built by a prior SuperUROP student, Emily Skilling, in 2020, and had been used by my graduate student advisor for some experiments since then. When I joined the project I was tasked with finding some way to quantify static balance ability using the instrumented cane, beyond the data collection which had already been performed. My main contribution–and the focus of my work during SuperUROP–was the development of a method to predict postural sway velocity, which is a metric that provides insight into balance health and fall risk in older adults.

I began by collecting validation data from sixteen young adults, and I designed a feature computation and selection pipeline, which ultimately made a prediction of a subject’s sway velocity for a window of data using linear regression. After finding success with the validation subjects, we moved on to working with eight participants over 65.

Burns offers a closer look at the cane, which can help measure movements that would be imperceptible to watching humans.

Your SuperUROP project isn’t the only way you’ve explored assistive technology, correct–it’s another outgrowth of your primary research interest.

Yes! I helped restart the MIT Assistive Technology Club, where we pair small teams of MIT students with a person who has a disability, who works with them as a codesigner to develop a prototype that improves their daily life. The teams meet with their codesigner every week, to ensure that the project is guided by the needs of the person who will use the device or software. I worked as the club’s project manager, and met with leads from our six teams each week to provide guidance in technical and logistical aspects of their projects, which are really diverse.

Learn more about the MIT Assistive Technology Club

A robotic arm for meals

One team was developing robotic self-feeding devices for two people with limited mobility from the neck down. They designed and built a robotic arm, which featured a custom utensil at the end for effectively dispensing a variety of food in a safe way. This project will be continuing this fall!

A LLM-powered visual aid

Our second team worked on an app for people who are blind, which delivers the most important information about the scene in front of them using better LLM prompting and computer vision. The framework for this project is now finished, and will be finishing up this summer!

An accommodation explainer

A third team designed an app that a person with a disability could use to design a page that has information about their accommodations, which could be accessed through an NFC tag attached to their wheelchair. The goal was to provide a faster way for new people to understand a person’s needs, without the person having to verbally explain over and over when they would prefer not to.

Mods for wheelchairs

The fourth team was working with a person who uses a wheelchair, and is interested in producing an open-source toolkit for improving hospital-style wheelchairs in low-income communities. The hospital-style wheelchairs commonly used around the world are cheaply made, and due to their design can result in serious stress injuries when a person uses them for long periods of time. This team was focused on making simple modifications to wheelchairs that use locally available materials and machines, and will be continuing their work next semester.

A better game controller

Another team was designing a one-handed video game controller with a woman who has partial paralysis. This project is almost complete, and has focused on PCB design, and making a working controller which is comfortable to use for long periods of time.

Memory assistance for faces

Our last team built a memory assistance app with a codesigner who struggled to remember faces. This app was designed to help the user quickly narrow down what a person’s name was based on characteristics and the circumstances that they were interacting with the user.

Tell me about a problem or roadblock you encountered during your project, and how you solved it.

Human subject data is inherently noisy and balance is very complex, which makes learning behavior with limited amounts of data difficult. I struggled initially in my project to make any sense of the data, and tried to find consistent trends through experimenting with different balance metrics with some success. I ultimately found that machine learning methods are incredibly good at extracting information that would be impossible for a person to see. Although I did not have prior experience, I dedicated time to learning and becoming comfortable with data processing and machine learning. I also read through previous journal articles to better understand how to extract useful features in human balance.

If you were to develop your cane project even further, what would be your next steps?

I think that with a combination of static balance assessment and gait analysis, you could construct an even stronger prediction of a person’s balance health and future fall risk which would be of interest to physicians. Also, the cane is intended to be used in everyday life and provide continuous assessment, so I would like to organize a much larger study where participants use the instrumented cane over the course of a month. Most falls occur in the home, so I think it would be particularly interesting to learn about balance in activities of daily living through the take-home walking cane.

What surprised you about lab work, or was different from your expectations? What do you think younger students (say, in high school) should know about doing STEM research?

I think one thing that surprised me about research (and invention in general) is how much is built upon the work of people who came before. When I was in high school, I would often try to “figure things out” without help from other resources, because when you’re in school working on homework, Googling questions is considered cheating! The biggest part of research that I came to understand is that an individual researcher is just one part of a global community, who are working together to achieve amazing things. Each paper or discovery is one piece in a large picture, which eventually leads to a major jump forwards. When you think about being a scientist or an engineer, it’s crucial to remember that you can and should rely on your collaborators, and the thousands of hours of work others have already done.

Tell me about the experience of presenting your research at the SuperUROP poster session. What did preparing your poster and talking with outside folks teach you about scientific communication?

The poster session was a great way to collect and organize the story of my project, and speaking to people from outside of my research was an excellent way to generate ideas going forward. I had the chance to speak with multiple people about their personal experience with balance issues; either their own, or of their family. My advisors in SuperUROP really helped me develop my poster, and getting direct constructive feedback on the aspects of both my elevator pitch and poster design was extremely helpful for improving my ability as a presenter. The poster session itself was also good practice, repeating a pitch for an hour is a great way to identify what is and isn’t important about my work!

“An individual researcher is just one part of a global community, who are working together to achieve amazing things.”

Max Burns

SuperUROP participant

Did your SuperUROP experience affect how you thought about possible future career paths?

Yes, SuperUROP helped to confirm that I want to continue academic research as a graduate student this fall! I also aim to be a professor someday, and the seminars held in class helped to provide a weekly perspective of the paths that MIT professors have taken to succeed. It was fun to hear from so many different people over the past year, and I learned a lot about the technical aspects of their work too! I think it’s incredibly helpful to learn more about cutting edge technology for any field.

Helping nonexperts build advanced generative AI models

Posted on June 28, 2024 by Jane Halpern - News

The impact of artificial intelligence will never be equitable if there’s only one company that builds and controls the models (not to mention the data that go into them). Unfortunately, today’s AI models are made up of billions of parameters that must be trained and tuned to maximize performance for each use case, putting the most powerful AI models out of reach for most people and companies.

MosaicML started with a mission to make those models more accessible. The company, which counts Jonathan Frankle PhD ’23 and MIT Associate Professor Michael Carbin as co-founders, developed a platform that let users train, improve, and monitor open-source models using their own data. The company also built its own open-source models using graphical processing units (GPUs) from Nvidia.

The approach made deep learning, a nascent field when MosaicML first began, accessible to far more organizations as excitement around generative AI and large language models (LLMs) exploded following the release of Chat GPT-3.5. It also made MosaicML a powerful complementary tool for data management companies that were also committed to helping organizations make use of their data without giving it to AI companies.

Last year, that reasoning led to the acquisition of MosaicML by Databricks, a global data storage, analytics, and AI company that works with some of the largest organizations in the world. Since the acquisition, the combined companies have released one of the highest performing open-source, general-purpose LLMs yet built. Known as DBRX, this model has set new benchmarks in tasks like reading comprehension, general knowledge questions, and logic puzzles.

Since then, DBRX has gained a reputation for being one of the fastest open-source LLMs available and has proven especially useful at large enterprises.

More than the model, though, Frankle says DBRX is significant because it was built using Databricks tools, meaning any of the company’s customers can achieve similar performance with their own models, which will accelerate the impact of generative AI.

“Honestly, it’s just exciting to see the community doing cool things with it,” Frankle says. “For me as a scientist, that’s the best part. It’s not the model, it’s all the amazing stuff the community is doing on top of it. That’s where the magic happens.”

Making algorithms efficient

Frankle earned bachelor’s and master’s degrees in computer science at Princeton University before coming to MIT to pursue his PhD in 2016. Early on at MIT, he wasn’t sure what area of computing he wanted to study. His eventual choice would change the course of his life.

Frankle ultimately decided to focus on a form of artificial intelligence known as deep learning. At the time, deep learning and artificial intelligence did not inspire the same broad excitement as they do today. Deep learning was a decades-old area of study that had yet to bear much fruit.

“I don’t think anyone at the time anticipated deep learning was going to blow up in the way that it did,” Frankle says. “People in the know thought it was a really neat area and there were a lot of unsolved problems, but phrases like large language model (LLM) and generative AI weren’t really used at that time. It was early days.”

Things began to get interesting with the 2017 release of a now-infamous paper by Google researchers, in which they showed a new deep-learning architecture known as the transformer was surprisingly effective as language translation and held promise across a number of other applications, including content generation.

In 2020, eventual Mosaic co-founder and tech executive Naveen Rao emailed Frankle and Carbin out of the blue. Rao had read a paper the two had co-authored, in which the researchers showed a way to shrink deep-learning models without sacrificing performance. Rao pitched the pair on starting a company. They were joined by Hanlin Tang, who had worked with Rao on a previous AI startup that had been acquired by Intel.

The founders started by reading up on different techniques used to speed up the training of AI models, eventually combining several of them to show they could train a model to perform image classification four times faster than what had been achieved before.

“The trick was that there was no trick,” Frankle says. “I think we had to make 17 different changes to how we trained the model in order to figure that out. It was just a little bit here and a little bit there, but it turns out that was enough to get incredible speed-ups. That’s really been the story of Mosaic.”

The team showed their techniques could make models more efficient, and they released an open-source large language model in 2023 along with an open-source library of their methods. They also developed visualization tools to let developers map out different experimental options for training and running models.

MIT’s E14 Fund invested in Mosaic’s Series A funding round, and Frankle says E14’s team offered helpful guidance early on. Mosaic’s progress enabled a new class of companies to train their own generative AI models.

“There was a democratization and an open-source angle to Mosaic’s mission,” Frankle says. “That’s something that has always been very close to my heart. Ever since I was a PhD student and had no GPUs because I wasn’t in a machine learning lab and all my friends had GPUs. I still feel that way. Why can’t we all participate? Why can’t we all get to do this stuff and get to do science?”

Open sourcing innovation

Databricks had also been working to give its customers access to AI models. The company finalized its acquisition of MosaicML in 2023 for a reported $1.3 billion.

“At Databricks, we saw a founding team of academics just like us,” Frankle says. “We also saw a team of scientists who understand technology. Databricks has the data, we have the machine learning. You can’t do one without the other, and vice versa. It just ended up being a really good match.”

In March, Databricks released DBRX, which gave the open-source community and enterprises building their own LLMs capabilities that were previously limited to closed models.

“The thing that DBRX showed is you can build the best open-source LLM in the world with Databricks,” Frankle says. “If you’re an enterprise, the sky’s the limit today.”

Frankle says Databricks’ team has been encouraged by using DBRX internally across a wide variety of tasks.

“It’s already great, and with a little fine-tuning it’s better than the closed models,” he says. “You’re not going be better than GPT for everything. That’s not how this works. But nobody wants to solve every problem. Everybody wants to solve one problem. And we can customize this model to make it really great for specific scenarios.”

As Databricks continues pushing the frontiers of AI, and as competitors continue to invest huge sums into AI more broadly, Frankle hopes the industry comes to see open source as the best path forward.

“I’m a believer in science and I’m a believer in progress and I’m excited that we’re doing such exciting science as a field right now,” Frankle says. “I’m also a believer in openness, and I hope that everybody else embraces openness the way we have. That’s how we got here, through good science and good sharing.”

Study reveals why AI models that analyze medical images can be biased

Posted on June 28, 2024 by Jane Halpern - News

Artificial intelligence models often play a role in medical diagnoses, especially when it comes to analyzing images such as X-rays. However, studies have found that these models don’t always perform well across all demographic groups, usually faring worse on women and people of color.

These models have also been shown to develop some surprising abilities. In 2022, MIT researchers reported that AI models can make accurate predictions about a patient’s race from their chest X-rays — something that the most skilled radiologists can’t do.

That research team has now found that the models that are most accurate at making demographic predictions also show the biggest “fairness gaps” — that is, discrepancies in their ability to accurately diagnose images of people of different races or genders. The findings suggest that these models may be using “demographic shortcuts” when making their diagnostic evaluations, which lead to incorrect results for women, Black people, and other groups, the researchers say.

“It’s well-established that high-capacity machine-learning models are good predictors of human demographics such as self-reported race or sex or age. This paper re-demonstrates that capacity, and then links that capacity to the lack of performance across different groups, which has never been done,” says Marzyeh Ghassemi, an MIT associate professor of electrical engineering and computer science, a member of MIT’s Institute for Medical Engineering and Science, and the senior author of the study.

The researchers also found that they could retrain the models in a way that improves their fairness. However, their approached to “debiasing” worked best when the models were tested on the same types of patients they were trained on, such as patients from the same hospital. When these models were applied to patients from different hospitals, the fairness gaps reappeared.

“I think the main takeaways are, first, you should thoroughly evaluate any external models on your own data because any fairness guarantees that model developers provide on their training data may not transfer to your population. Second, whenever sufficient data is available, you should train models on your own data,” says Haoran Zhang, an MIT graduate student and one of the lead authors of the new paper. MIT graduate student Yuzhe Yang is also a lead author of the paper, which appears today in Nature Medicine. Judy Gichoya, an associate professor of radiology and imaging sciences at Emory University School of Medicine, and Dina Katabi, the Thuan and Nicole Pham Professor of Electrical Engineering and Computer Science at MIT, are also authors of the paper.

Removing bias

As of May 2024, the FDA has approved 882 AI-enabled medical devices, with 671 of them designed to be used in radiology. Since 2022, when Ghassemi and her colleagues showed that these diagnostic models can accurately predict race, they and other researchers have shown that such models are also very good at predicting gender and age, even though the models are not trained on those tasks.

“Many popular machine learning models have superhuman demographic prediction capacity — radiologists cannot detect self-reported race from a chest X-ray,” Ghassemi says. “These are models that are good at predicting disease, but during training are learning to predict other things that may not be desirable.”

In this study, the researchers set out to explore why these models don’t work as well for certain groups. In particular, they wanted to see if the models were using demographic shortcuts to make predictions that ended up being less accurate for some groups. These shortcuts can arise in AI models when they use demographic attributes to determine whether a medical condition is present, instead of relying on other features of the images.

Using publicly available chest X-ray datasets from Beth Israel Deaconess Medical Center in Boston, the researchers trained models to predict whether patients had one of three different medical conditions: fluid buildup in the lungs, collapsed lung, or enlargement of the heart. Then, they tested the models on X-rays that were held out from the training data.

Overall, the models performed well, but most of them displayed “fairness gaps” — that is, discrepancies between accuracy rates for men and women, and for white and Black patients.

The models were also able to predict the gender, race, and age of the X-ray subjects. Additionally, there was a significant correlation between each model’s accuracy in making demographic predictions and the size of its fairness gap. This suggests that the models may be using demographic categorizations as a shortcut to make their disease predictions.

The researchers then tried to reduce the fairness gaps using two types of strategies. For one set of models, they trained them to optimize “subgroup robustness,” meaning that the models are rewarded for having better performance on the subgroup for which they have the worst performance, and penalized if their error rate for one group is higher than the others.

In another set of models, the researchers forced them to remove any demographic information from the images, using “group adversarial” approaches. Both strategies worked fairly well, the researchers found.

“For in-distribution data, you can use existing state-of-the-art methods to reduce fairness gaps without making significant trade-offs in overall performance,” Ghassemi says. “Subgroup robustness methods force models to be sensitive to mispredicting a specific group, and group adversarial methods try to remove group information completely.”

Not always fairer

However, those approaches only worked when the models were tested on data from the same types of patients that they were trained on — for example, only patients from the Beth Israel Deaconess Medical Center dataset.

When the researchers tested the models that had been “debiased” using the BIDMC data to analyze patients from five other hospital datasets, they found that the models’ overall accuracy remained high, but some of them exhibited large fairness gaps.

“If you debias the model in one set of patients, that fairness does not necessarily hold as you move to a new set of patients from a different hospital in a different location,” Zhang says.

This is worrisome because in many cases, hospitals use models that have been developed on data from other hospitals, especially in cases where an off-the-shelf model is purchased, the researchers say.

“We found that even state-of-the-art models which are optimally performant in data similar to their training sets are not optimal — that is, they do not make the best trade-off between overall and subgroup performance — in novel settings,” Ghassemi says. “Unfortunately, this is actually how a model is likely to be deployed. Most models are trained and validated with data from one hospital, or one source, and then deployed widely.”

The researchers found that the models that were debiased using group adversarial approaches showed slightly more fairness when tested on new patient groups than those debiased with subgroup robustness methods. They now plan to try to develop and test additional methods to see if they can create models that do a better job of making fair predictions on new datasets.

The findings suggest that hospitals that use these types of AI models should evaluate them on their own patient population before beginning to use them, to make sure they aren’t giving inaccurate results for certain groups.

The research was funded by a Google Research Scholar Award, the Robert Wood Johnson Foundation Harold Amos Medical Faculty Development Program, RSNA Health Disparities, the Lacuna Fund, the Gordon and Betty Moore Foundation, the National Institute of Biomedical Imaging and Bioengineering, and the National Heart, Lung, and Blood Institute.

Startup aims to transform the power grid with superconducting transmission lines

Posted on June 28, 2024 by Jane Halpern - News

Last year in Woburn, Massachusetts, a power line was deployed across a 100-foot stretch of land. Passersby wouldn’t have found much interesting about the installation: The line was supported by standard utility poles, the likes of which most of us have driven by millions of times. In fact, the familiarity of the sight is a key part of the technology’s promise.

The lines are designed to transport five to 10 times the amount of power of conventional transmission lines, using essentially the same footprint and voltage level. That will be key to helping them overcome the regulatory hurdles and community opposition that has made increasing transmission capacity nearly impossible across large swaths of the globe, particularly in America and Europe, where new power distribution systems play a vital role in the shift to renewable energy and the resilience of the grid.

The lines are the product of years of work by the startup VEIR, which was co-founded by Tim Heidel ’05, SM ’06, SM ’09, PhD ’10. They make use of superconducting cables and a proprietary cooling system that will enable initial transmission capacity up to 400 megawatts and, in future versions, up to several gigawatts.

“We can deploy much higher power levels at much lower voltage, and so we can deploy the same high power but with a footprint and visual impact that is far less intrusive, and therefore can overcome a lot of the public opposition as well as siting and permitting barriers,” Heidel says.

VEIR’s solution comes at a time when more than 10,000 renewable energy projects at various stages of development are seeking permission to connect to U.S. grids. The White House has said the U.S. must more than double existing regional transmission capacity in order to reach 2035 decarbonization goals.

All of this comes as electricity demand is skyrocketing amid the increasing use of data centers and AI, and the electrification of everything from passenger vehicles to home heating systems.

Despite those trends, building high-power transmission lines remains stubbornly difficult.

“Building high-power transmission infrastructure can take a decade or more, and there’s been quite a few examples of projects that folks have had to abandon because they realize that there’s just so much opposition, or there’s too much complexity to pull it off cost effectively,” Heidel says. “We can drop down in voltage but carry the same amount of power because we can build systems that operate at much higher current levels, and that’s how our lines are able to melt into the background and avoid the same opposition.”

Heidel says VEIR has built a pipeline of interested customers including utilities, data center operators, industrial companies, and renewable energy developers. VEIR is aiming to complete its first commercial-scale pilot carrying high power in 2026.

The core innovation at VEIR is the cooling system, which is passively cooled with nitrogen. Photo: Courtesy of VEIR

A career in energy

Over more than a decade at MIT, Heidel went from learning about the fundamentals of electrical engineering to studying the electric grid and the power sector more broadly. That journey included earning a bachelor’s, master’s, and PhD from MIT’s Department of Electrical Engineering and Computer Science as well as a master’s in MIT’s Technology and Policy Program, which he earned while working toward his PhD.

“I got the energy bug and started to focus exclusively on energy and climate in graduate school,” Heidel says.

Following his PhD, Heidel was named research director of MIT’s Future of the Electric Grid study, which was completed in 2011.

“That was a fantastic opportunity at the outset of my career to survey the entire landscape and understand challenges facing the power grid and the power sector more broadly,” Heidel says. “It gave me a good foundation for understanding the grid, how it works, who’s involved, how decisions get made, how expansion works, and it looked out over the next 30 years.”

After leaving MIT, Heidel worked at the Department of Energy’s Advanced Research Projects Agency-Energy (ARPA-E) and then at Bill Gates’ Breakthrough Energy Ventures (BEV) investment firm, where he continued studying transmission.

“Just about every single decarbonization scenario and study that’s been published in the last two decades concludes that to achieve aggressive greenhouse gas emissions reductions, we’re going to have to double or triple the scale of power grids around the world,” Heidel says. “But when we looked at the data on how fast grids were being expanded, the ease with which transmission lines could be built, the cost of building new transmission, just about every indicator was heading in the wrong direction. Transmission was getting more expensive over time and taking longer to build. We desperately need to find a new solution.”

Unlike traditional transmission lines made from steel and aluminum, VEIR’s transmission lines leverage decades of progress in the development of high-temperature superconducting tapes and other materials. Some of that progress has been driven by the nuclear fusion industry, which incorporates superconducting materials into some of their nuclear reactor designs.

But the core innovation at VEIR is the cooling system. VEIR co-founder and advisor Steve Ashworth developed the rough idea for the cooling system more than 15 years ago at Los Alamos National Laboratory as part of a larger Department of Energy-funded research project. When the project was shut down, the idea was largely forgotten.

Heidel and others at Breakthrough Energy Ventures became aware of the innovation in 2019 while researching transmission. Today VEIR’s system is passively cooled with nitrogen, which runs through a vacuum-insulated pipe that surrounds a superconducting cable. Heat exchange units are also used on some transmission towers.

Heidel says transmission lines designed to carry that much power are typically far bigger than VEIR’s design, and other attempts at shrinking the footprint of high-power lines were limited to short distances underground.

“High power requires high voltage, and high voltage requires tall towers and wide right of ways, and those tall towers and those wide right of ways are deeply unpopular,” Heidel says. “That is a universal truth across just about the entire world.”

Moving power around the world

VEIR’s first alternating current (AC) overhead product line is capable of transmission capacities up to 400 megawatts and voltages of up to 69 kilovolts, and the company plans to scale to higher voltage and higher-power products in the future, including direct current (DC) lines.

VEIR will sell its equipment to the companies installing transmission lines, with a primary focus on the U.S. market.

In the longer term, Heidel believes VEIR’s technology is needed as soon as possible to meet rising electricity demands and new renewable energy projects around the globe.

Wireless receiver blocks interference for better mobile device performance

Posted on June 28, 2024 by Jane Halpern - News

The growing prevalence of high-speed wireless communication devices, from 5G mobile phones to sensors for autonomous vehicles, is leading to increasingly crowded airwaves. This makes the ability to block interfering signals that can hamper device performance an even more important — and more challenging — problem.

With these and other emerging applications in mind, MIT researchers demonstrated a new millimeter-wave multiple-input-multiple-output (MIMO) wireless receiver architecture that can handle stronger spatial interference than previous designs. MIMO systems have multiple antennas, enabling them to transmit and receive signals from different directions. Their wireless receiver senses and blocks spatial interference at the earliest opportunity, before unwanted signals have been amplified, which improves performance.

Key to this MIMO receiver architecture is a special circuit that can target and cancel out unwanted signals, known as a nonreciprocal phase shifter. By making a novel phase shifter structure that is reconfigurable, low-power, and compact, the researchers show how it can be used to cancel out interference earlier in the receiver chain.

Their receiver can block up to four times more interference than some similar devices. In addition, the interference-blocking components can be switched on and off as needed to conserve energy.

In a mobile phone, such a receiver could help mitigate signal quality issues that can lead to slow and choppy Zoom calling or video streaming.

“There is already a lot of utilization happening in the frequency ranges we are trying to use for new 5G and 6G systems. So, anything new we are trying to add should already have these interference-mitigation systems installed. Here, we’ve shown that using a nonreciprocal phase shifter in this new architecture gives us better performance. This is quite significant, especially since we are using the same integrated platform as everyone else,” says Negar Reiskarimian, the X-Window Consortium Career Development Assistant Professor in the Department of Electrical Engineering and Computer Science (EECS), a member of the Microsystems Technology Laboratories and Research Laboratory of Electronics (RLE), and the senior author of a paper on this receiver.

Reiskarimian wrote the paper with EECS graduate students Shahabeddin Mohin, who is the lead author, Soroush Araei, and Mohammad Barzgari, an RLE postdoc. The work was recently presented at the IEEE Radio Frequency Circuits Symposium and received the Best Student Paper Award.

Blocking interference

Digital MIMO systems have an analog and a digital portion. The analog portion uses antennas to receive signals, which are amplified, down-converted, and passed through an analog-to-digital converter before being processed in the digital domain of the device. In this case, digital beamforming is required to retrieve the desired signal.

But if a strong, interfering signal coming from a different direction hits the receiver at the same time as a desired signal, it can saturate the amplifier so the desired signal is drowned out. Digital MIMOs can filter out unwanted signals, but this filtering occurs later in the receiver chain. If the interference is amplified along with the desired signal, it is more difficult to filter out later.

“The output of the initial low-noise amplifier is the first place you can do this filtering with minimal penalty, so that is exactly what we are doing with our approach,” Reiskarimian says.

The researchers built and installed four nonreciprocal phase shifters immediately at the output of the first amplifier in each receiver chain, all connected to the same node. These phase shifters can pass signal in both directions and sense the angle of an incoming interfering signal. The devices can adjust their phase until they cancel out the interference.

The phase of these devices can be precisely tuned, so they can sense and cancel an unwanted signal before it passes to the rest of the receiver, blocking interference before it affects any other parts of the receiver. In addition, the phase shifters can follow signals to continue blocking interference if it changes location.

“If you start getting disconnected or your signal quality goes down, you can turn this on and mitigate that interference on the fly. Because ours is a parallel approach, you can turn it on and off with minimal effect on the performance of the receiver itself,” Reiskarimian adds.

A compact device

In addition to making their novel phase shifter architecture tunable, the researchers designed them to use less space on the chip and consume less power than typical nonreciprocal phase shifters.

Once the researchers had done the analysis to show their idea would work, their biggest challenge was translating the theory into a circuit that achieved their performance goals. At the same time, the receiver had to meet strict size restrictions and a tight power budget, or it wouldn’t be useful in real-world devices.

In the end, the team demonstrated a compact MIMO architecture on a 3.2-square-millimeter chip that could block signals which were up to four times stronger than what other devices could handle. Simpler than typical designs, their phase shifter architecture is also more energy efficient.

Moving forward, the researchers want to scale up their device to larger systems, as well as enable it to perform in the new frequency ranges utilized by 6G wireless devices. These frequency ranges are prone to powerful interference from satellites. In addition, they would like to adapt nonreciprocal phase shifters to other applications.

This research was supported, in part, by the MIT Center for Integrated Circuits and Systems.

Advancing technology for aquaculture

Posted on June 27, 2024 by Jane Halpern - News

According to the National Oceanic and Atmospheric Administration, aquaculture in the United States represents a $1.5 billion industry annually. Like land-based farming, shellfish aquaculture requires healthy seed production in order to maintain a sustainable industry. Aquaculture hatchery production of shellfish larvae — seeds — requires close monitoring to track mortality rates and assess health from the earliest stages of life.

Careful observation is necessary to inform production scheduling, determine effects of naturally occurring harmful bacteria, and ensure sustainable seed production. This is an essential step for shellfish hatcheries but is currently a time-consuming manual process prone to human error.

With funding from MIT’s Abdul Latif Jameel Water and Food Systems Lab (J-WAFS), MIT Sea Grant is working with Associate Professor Otto Cordero of the MIT Department of Civil and Environmental Engineering, Professor Taskin Padir and Research Scientist Mark Zolotas at the Northeastern University Institute for Experiential Robotics, and others at the Aquaculture Research Corporation (A.R.C.), and the Cape Cod Commercial Fishermen’s Alliance, to advance technology for the aquaculture industry. Located on Cape Cod, A.R.C. is a leading shellfish hatchery, farm, and wholesaler that plays a vital role in providing high-quality shellfish seed to local and regional growers.

Two MIT students have joined the effort this semester, working with Robert Vincent, MIT Sea Grant’s assistant director of advisory services, through the Undergraduate Research Opportunities Program (UROP).

First-year student Unyime Usua and sophomore Santiago Borrego are using microscopy images of shellfish seed from A.R.C. to train machine learning algorithms that will help automate the identification and counting process. The resulting user-friendly image recognition tool aims to aid aquaculturists in differentiating and counting healthy, unhealthy, and dead shellfish larvae, improving accuracy and reducing time and effort.

Vincent explains that AI is a powerful tool for environmental science that enables researchers, industry, and resource managers to address challenges that have long been pinch points for accurate data collection, analysis, predictions, and streamlining processes. “Funding support from programs like J-WAFS enable us to tackle these problems head-on,” he says.

ARC faces challenges with manually quantifying larvae classes, an important step in their seed production process. “When larvae are in their growing stages they are constantly being sized and counted,” explains Cheryl James, A.R.C. larval/juvenile production manager. “This process is critical to encourage optimal growth and strengthen the population.”

Developing an automated identification and counting system will help to improve this step in the production process with time and cost benefits. “This is not an easy task,” says Vincent, “but with the guidance of Dr. Zolotas at the Northeastern University Institute for Experiential Robotics and the work of the UROP students, we have made solid progress.”

The UROP program benefits both researchers and students. Involving MIT UROP students in developing these types of systems provides insights into AI applications that they might not have considered, providing opportunities to explore, learn, and apply themselves while contributing to solving real challenges.

Borrego saw this project as an opportunity to apply what he’d learned in class 6.390 (Introduction to Machine Learning) to a real-world issue. “I was starting to form an idea of how computers can see images and extract information from them,” he says. “I wanted to keep exploring that.”

Usua decided to pursue the project because of the direct industry impacts it could have. “I’m pretty interested in seeing how we can utilize machine learning to make people’s lives easier. We are using AI to help biologists make this counting and identification process easier.” While Usua wasn’t familiar with aquaculture before starting this project, she explains, “Just hearing about the hatcheries that Dr. Vincent was telling us about, it was unfortunate that not a lot of people know what’s going on and the problems that they’re facing.”

On Cape Cod alone, aquaculture is an $18 million per year industry. But the Massachusetts Division of Marine Fisheries estimates that hatcheries are only able to meet 70–80 percent of seed demand annually, which impacts local growers and economies. Through this project, the partners aim to develop technology that will increase seed production, advance industry capabilities, and help understand and improve the hatchery microbiome.

Borrego explains the initial challenge of having limited data to work with. “Starting out, we had to go through and label all of the data, but going through that process helped me learn a lot.” In true MIT fashion, he shares his takeaway from the project: “Try to get the best out of what you’re given with the data you have to work with. You’re going to have to adapt and change your strategies depending on what you have.”

Usua describes her experience going through the research process, communicating in a team, and deciding what approaches to take. “Research is a difficult and long process, but there is a lot to gain from it because it teaches you to look for things on your own and find your own solutions to problems.”

In addition to increasing seed production and reducing the human labor required in the hatchery process, the collaborators expect this project to contribute to cost savings and technology integration to support one of the most underserved industries in the United States.

Borrego and Usua both plan to continue their work for a second semester with MIT Sea Grant. Borrego is interested in learning more about how technology can be used to protect the environment and wildlife. Usua says she hopes to explore more projects related to aquaculture. “It seems like there’s an infinite amount of ways to tackle these issues.”

EECS 2024 Awards

Posted on June 26, 2024 by Jane Halpern - EECS Celebrates Awards

Winners of annual awards given by the Department of EECS gather for a ceremony.

It’s the summer of 2024, and we want to celebrate the accomplishments and contributions of our incredible EECS community by sharing some of the awards given by the department this year. Congratulations to all the winners!