Artificial Intelligence

MIT researchers have developed a type of neural network that learns on the job, not just during its training phase. These flexible algorithms, dubbed “liquid” networks, change their underlying equations to continuously adapt to new data inputs. The advance could aid decision making based on data streams that change over time, including those involved in medical diagnosis and autonomous driving.
“This is a way forward for the future of robot control, natural language processing, video processing — any form of time series data processing,” says Ramin Hasani, the study’s lead author. “The potential is really significant.”
The research will be presented at February’s AAAI Conference on Artificial Intelligence. In addition to Hasani, a postdoc in the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL), MIT co-authors include Daniela Rus, CSAIL director and the Andrew and Erna Viterbi Professor of Electrical Engineering and Computer Science, and PhD student Alexander Amini. Other co-authors include Mathias Lechner of the Institute of Science and Technology Austria and Radu Grosu of the Vienna University of Technology.
Time series data are both ubiquitous and vital to our understanding the world, according to Hasani. “The real world is all about sequences. Even our perception — you’re not perceiving images, you’re perceiving sequences of images,” he says. “So, time series data actually create our reality.”
He points to video processing, financial data, and medical diagnostic applications as examples of time series that are central to society. The vicissitudes of these ever-changing data streams can be unpredictable. Yet analyzing these data in real time, and using them to anticipate future behavior, can boost the development of emerging technologies like self-driving cars. So Hasani built an algorithm fit for the task.
Hasani designed a neural network that can adapt to the variability of real-world systems. Neural networks are algorithms that recognize patterns by analyzing a set of “training” examples. They’re often said to mimic the processing pathways of the brain — Hasani drew inspiration directly from the microscopic nematode, C. elegans. “It only has 302 neurons in its nervous system,” he says, “yet it can generate unexpectedly complex dynamics.”
Hasani coded his neural network with careful attention to how C. elegans neurons activate and communicate with each other via electrical impulses. In the equations he used to structure his neural network, he allowed the parameters to change over time based on the results of a nested set of differential equations.
This flexibility is key. Most neural networks’ behavior is fixed after the training phase, which means they’re bad at adjusting to changes in the incoming data stream. Hasani says the fluidity of his “liquid” network makes it more resilient to unexpected or noisy data, like if heavy rain obscures the view of a camera on a self-driving car. “So, it’s more robust,” he says.
There’s another advantage of the network’s flexibility, he adds: “It’s more interpretable.”
Hasani says his liquid network skirts the inscrutability common to other neural networks. “Just changing the representation of a neuron,” which Hasani did with the differential equations, “you can really explore some degrees of complexity you couldn’t explore otherwise.” Thanks to Hasani’s small number of highly expressive neurons, it’s easier to peer into the “black box” of the network’s decision making and diagnose why the network made a certain characterization.
“The model itself is richer in terms of expressivity,” says Hasani. That could help engineers understand and improve the liquid network’s performance.
Hasani’s network excelled in a battery of tests. It edged out other state-of-the-art time series algorithms by a few percentage points in accurately predicting future values in datasets, ranging from atmospheric chemistry to traffic patterns. “In many applications, we see the performance is reliably high,” he says. Plus, the network’s small size meant it completed the tests without a steep computing cost. “Everyone talks about scaling up their network,” says Hasani. “We want to scale down, to have fewer but richer nodes.”
Hasani plans to keep improving the system and ready it for industrial application. “We have a provably more expressive neural network that is inspired by nature. But this is just the beginning of the process,” he says. “The obvious question is how do you extend this? We think this kind of network could be a key element of future intelligence systems.”
This research was funded, in part, by Boeing, the National Science Foundation, the Austrian Science Fund, and Electronic Components and Systems for European Leadership. More

Last fall, the MIT Stephen A. Schwarzman College of Computing embarked on a project to design and construct a new building on Vassar Street in Cambridge, at the former site of Building 44. Working with Skidmore, Owings & Merrill (SOM), the design for the new building is taking shape, with plans for the exterior façade now complete.
The proposed project will establish a home for the MIT Schwarzman College of Computing, providing state-of-the-art space for computing research and education. The building’s central location in the Vassar Street block between Main Street and Massachusetts Avenue will help form a new cluster of connectivity, and will enable the space to have a multifaceted role. The project has been reviewed extensively with city planning staff and will be presented to the Cambridge Planning Board for review and approval.
“The new building will serve as a hub for both disciplinary and interdisciplinary work in computing and collaboration at MIT. It will also contain inviting, community-oriented spaces where we can bring a mix of people together,” says Daniel Huttenlocher, dean of the MIT Schwarzman College of Computing. “The middle floors of the building are designed for research groups and other parts of the college, while the lower floors and an event space on the top floor are oriented towards the MIT community and others more broadly.”
As seen in architectural renderings, the building’s transparent and outward-looking design will give passersby a direct view into the first two floors. These floors will offer multiple convening areas for the campus community and the public to come in and engage with the college, including a 250-seat lecture hall, a suite of student spaces for project-based work and tutoring, a café, and a variety of seating options for studying and social interactions.
The exterior’s shingled-glass façade creates a visual cue that differentiates the lower floors from the upper section. Research space will be concentrated on four floors, to house 50 new and existing faculty members working in computing and their research groups, including faculty from schools across MIT. Collaborative research spaces will be spread throughout those floors, including gathering areas that facilitate spontaneous interactions.
The building will also support other college activities such as the MIT Quest for Intelligence, and will include space for visitors and visiting scholars, as well as administrative areas. In addition, spaces in the building will accommodate cross-cutting areas of the college, currently the Social and Ethical Responsibilities of Computing and the Common Ground for Computing Education.
An event space and an outdoor terrace are planned for the top floor of the building, which will offer views of the entire MIT campus, into Boston’s Back Bay, and portions of the Boston skyline.
Throughout the design process, the project team has made access and sustainability priorities and is aiming for a minimum of Leadership in Energy and Environmental Design (LEED) Gold certification for building and construction. Toward that goal, the south-facing side of the building will feature a double-skin façade constructed from state-of-the-art interlocking glass units that create a deep sealed cavity — a design solution that is expected to reduce energy consumption by approximately 27 percent over baseline double glazing units, providing greater improvement than a typical façade.
“It is part of a holistic vision for sustainability,” says Colin Koop, SOM design partner. “We will monitor the entire process of creating this building with embodied carbon tracking, and when the building is complete, it will include on-site stormwater management, resiliency against flooding, a large green roof, efficient heating and cooling, and more.” Other features and goals will include elevators placed by the main entrance to increase visibility and accessibility to patrons entering the building and fixtures that will reduce indoor potable water usage by 36.7 percent from the Environmental Protection Agency baseline.
The overall design process is slated to wrap up this coming fall. To prepare for the start of full construction, which is anticipated to begin this summer, several research and education teams occupying Building 44 were relocated within existing campus buildings last year and have been operating successfully from those locations since. In addition, enabling work on the site was initiated last spring. The project now is in the final stages of preparation, with the removal of Building 44 having taken place last September and relocation of utilities almost done. The building is planned for completion by summer 2023. More

Between remote learning, more time spent at home, and working parents trying to keep their kids occupied, children across the United States have clocked in record-breaking hours of screen time during the pandemic. Much of it is supervised and curated by teachers or parents — but increasingly, kids of all ages are watching videos, playing games, and interacting with devices powered by artificial intelligence. As head of the Personal Robots group and AI Education at MIT, Media Lab Professor Cynthia Breazeal is on a mission to help this generation of young people to grow up understanding the AI they use.
At “AI Education: Research and Practice,” an Open Learning Talks event in December, Breazeal shared her vision for educating students not only about how AI works, but how to design and use it themselves — an initiative she calls AI Literacy for All. The AI Education project Breazeal is leading at MIT is a collaboration between MIT Open Learning and the Abdul Latif Jameel World Education Lab, the Media Lab, and the MIT Schwarzman College of Computing. Through research projects, hands-on activities, and scalable learning modules, Breazeal and her AI Education affiliates across MIT are creating a robust resource hub for educators, parents, and learners of all ages to understand how AI functions in different day-to-day roles, and how to approach both using and creating artificial intelligence with a basis in ethics, inclusion, and empathy.

Open Learning Talks | AI Education: Research and Practice

“It’s at this intersection of human psychology, engagement, and AI and technology, and we’re learning a lot,” Breazeal said as she explained her group’s research to the audience. “We’re not trying to build technologies to replace teachers or compete with parents. These are fluffy, pet-like robots, but they can engage children in this interaction where there are aspects like a motivating ally, like a friend … there are aspects like this companion animal, and this nonjudgmental companion animal gives the nature of this relationship this very different flavor, where even if they’re embarrassed to make mistakes in front of their teacher or their friends, they seem not to be in front of the robot — and you can’t learn if you’re not willing to take learning risks.” 
Breazeal shared examples from her Personal Robots group’s efforts, including recent studies on personalized learning companions for early childhood education, developing comprehensive K-12 AI literacy programs, and creating tools to help kids get creative using AI technologies.
“So how do you empower kids to create things with AI? You’re not going to put a middle-schooler on Tensorflow and say ‘Good luck,’ right?” Breazeal said. “MIT is the home base for things like Scratch and App Inventor, so the team is taking these more advanced AI methods and curricula and concepts, and augmenting these platforms to empower kids to use these AI technologies, to learn about them and then design projects of their own, and port them to different kinds of platforms.”  
Host Professor Eric Klopfer, director of the Scheller Teacher Education Program and the Education Arcade at MIT and head of MIT Comparative Media Studies and Writing, engaged Breazeal in a dialogue about all aspects of AI education and fielded questions from the live audience, ranging from emotional connection with robots to screen time, data collection, and representation in research and design. 
“How does AI in education narrow the gap that we see between socioeconomic groups? How do we see AI bridging that gap rather than widening the gap?” asked Klopfer, as he and Breazeal shared insights on training teachers, providing hands-on activities and paper prototyping to expand access and inclusion on technology education. “The technology itself is not the impetus for the divide anymore; it’s the way the technologies are being used, and the way people are trained to be able to use them,” Klopfer said. “It’s so key that we don’t repeat our mistakes from past technological innovations, where we just distribute devices to schools without thinking about the training and expertise that needs to go with that.”
And in an increasingly tech-driven society, access and education are key to creating equity for, and encouraging thoughtful participation from, all users. “We want a much more diverse, inclusive group of people being able to participate in shaping this future [with AI],” said Breazeal.
Launched last fall, Open Learning Talks is a public, online event series that features conversations between leaders from MIT and around the world, sharing their research and insights on education, teaching, and the science of learning. Upcoming events include William Bonvillian and Sanjay Sarma discussing their new book, “Workforce Education,” on Feb. 23; and Professor D. Fox Harrell and Rocky Bucano, executive director of the Universal Hip Hop Museum, in mid-March.  More

As part of the MIT Task Force on the Work of the Future’s series of research briefs, Professor Thomas Malone, Professor Daniela Rus, and Robert Laubacher collaborated on “Artificial Intelligence and the Future of Work,” a brief that provides a comprehensive overview of AI today and what lies at the AI frontier. The authors delve into the question of how work will change with AI and provide policy prescriptions that speak to different parts of society. Thomas Malone is director of the MIT Center for Collective Intelligence and the Patrick J. McGovern Professor of Management in the MIT Sloan School of Management. Daniela Rus is director of the Computer Science and Artificial Intelligence Laboratory, the Andrew and Erna Viterbi Professor of Electrical Engineering and Computer Science, and a member of the MIT Task Force on the Work of the Future. Robert Laubacher is associate director of the MIT Center for Collective Intelligence.Here, Malone and Rus provide an overview of their research.Q: You argue in your brief that despite major recent advances, artificial intelligence is nowhere close to matching the breadth and depth of perception, reasoning, communication, and creativity of people. Could you explain some of the limitations of AI?Rus: Despite recent and significant strides in the AI field, and great promise for the future, today’s AI systems are still quite limited in their ability to reason, make decisions, interact reliably with people and the physical world. Some of today’s greatest successes are due to a machine learning method called deep learning. These systems are trained using vast amounts of data that needs to be manually labeled. Their performance is dependent on the quantity and quality of data used to train them. The larger the training set for the network, the better its performance, and, in turn, the better the product that relies on the machine learning engine. But training large models has high computation cost. Also, bad training data leads to bad performance: when the data has bias, the system response propagates the same bias.Another limitation of current AI systems is robustness. Current state-of-the-art classifiers achieve impressive performance on benchmarks, but their predictions tend to be brittle. Specifically, inputs that were initially classified correctly can become misclassified once a carefully constructed but indiscernible perturbation is added to them. An important consequence of the lack of robustness is the lack of trust. One of the worrisome factors about the use of AI is the lack of guarantee that an input will be processed and classified correctly. The complex nature of training and using neural networks leads to systems that are difficult for people to understand. The systems are not able to provide explanations for how they reached decisions.Q: What are the ways AI is complementing, or could complement, human work?Malone: Today’s AI programs have only specialized intelligence; they’re only capable of doing certain specialized tasks. But humans have a kind of general intelligence that lets them do a much broader range of things.That means some of the best ways for AI systems to complement human work is to do specialized tasks that computers can do better, faster, or more cheaply than people can. For example, AI systems can be helpful by doing tasks such as interpreting medical X-rays, evaluating the risk of fraud in a credit card charge, or generating unusual new product designs.And humans can use their social skills, common sense, and other kinds of general intelligence to do things computers can’t do well. For instance, people can provide emotional support to patients diagnosed with cancer. They can decide when to believe customer explanations for unusual credit card transactions, and they can reject new product designs that customers would probably never want.In other words, many of the most important uses of computers in the future won’t be replacing people; they’ll be working with people in human-computer groups — “superminds” — that can do things better than either people or computers alone could do.The possibilities here go far beyond what people usually think of when they hear a phrase like “humans in the loop,” Instead of AI technologies just being tools to augment individual humans, we believe that many of their most important uses will occur in the context of groups of humans — often connected by the internet. So we should move from thinking about humans in the loop to computers in the group.Q: What are some of your recommendations for education, business, and government regarding policies to help smooth the transition of AI technology adoption? Rus: In our report, we highlight four types of actions that can reduce the pain associated with job transitions: education and training, matching jobs to job seekers, creating new jobs, and providing counseling and financial support to people as they transition from old to new jobs. Importantly, we will need partnership among a broad range of institutions to get this work done.Malone: We expect that — as with all previous labor-saving technologies — AI will eventually lead to the creation of more new jobs than it eliminates. But we see many opportunities for different parts of society to help smooth this transition, especially for the individuals whose old jobs are disrupted and who cannot easily find new ones.For example, we believe that businesses should focus on applying AI in ways that don’t just replace people but that create new jobs by providing novel kinds of products and services. We recommend that all schools include computer literacy and computational thinking in their curricula, and we believe that community colleges should offer more reskilling and online micro-degree programs, often including apprenticeships at local employers.We think that current worker organizations (such as labor unions and professional associations) or new ones (perhaps called “guilds”) should expand their roles to provide benefits previously tied to formal employment (such as insurance and pensions, career development, social connections, a sense of identity, and income security).And we believe that governments should increase their investments in education and reskilling programs to make the American workforce once again the best-educated in the world. And they should reshape the legal and regulatory framework that governs work to encourage creating more new jobs. More

Pavement deterioration takes many forms. It can manifest in almost imperceptible flaws, like surface roughness, to much more evident distresses, such as web-like alligator cracks. While the causes of these distresses are numerous, one cause, in particular, can impose an intractable burden: the weight of a vehicle.
In a prize-winning paper, Fengdi Guo, a PhD candidate at the MIT Concrete Sustainability Hub, helps clarify the layered relationship between traffic weight and pavement deterioration. The machine learning models he proposes have found that traffic weight induces specific kinds of damage in asphalt pavements, accelerating their deterioration rates. Concrete pavements, however, proved insensitive to traffic weight.
The paper, “Assessing the Influence of Overweight Vehicles on Pavement Performance,” was awarded first place in The Aramis López Challenge Category of the LTPP Analysis Student Contest, a joint effort of the Federal Highway Administration (FHWA) and the American Society of Civil Engineers’ Transportation and Development Institute. Guo will present his findings at the 2021 Transportation Research Board annual meeting.
As one might expect, predicting pavement deterioration is crucial to maintaining road networks. And traffic — specifically, accumulative traffic weights during a period — can play a key role in how quickly a pavement deteriorates.
“The accumulative traffic weight is the product of two components: traffic volume, represented by the annual average daily truck traffic (AADTT), and traffic weight, represented by the approximate weight of a flatbed truck,” explains Guo. “If, for instance, AADTT on a segment were to increase by 1,000, the time between maintenances would decrease by five months, on average.”
When one considers the latest transportation trends, truck traffic weight is likely to become especially problematic; according to the U.S. Energy Information Administration, heavy- and medium-duty vehicle traffic is expected to grow by nearly 40 percent by 2050, far outstripping growth in passenger vehicle traffic.
Accommodating such heavy truck traffic will require more sophisticated tools — particularly because the relationship between traffic weight and pavement deterioration has remained uncharted.
“Though greater traffic weight indisputably deteriorates asphalt pavements,” says Guo, “the types of deterioration it causes are much more unclear. Numerous factors, from precipitation rates to the thickness of a single layer of a pavement, can alter how a pavement responds to the weight of a vehicle.”
To account for these many factors, researchers and engineers have tended to use either complex mechanistic models or data-driven models. The former focus narrowly on the mechanical properties of pavements, require large computational resources, and are not suitable for analyzing a pavement network. The latter can be applied to a pavement network, yet they cannot incorporate a pavement’s unique maintenance and deterioration history.
In his paper, Guo sought to expand the scope of data-driven models. Instead of simply estimating a pavement’s key historical factors, he incorporated them directly into his calculations.
His approach relies on what is known as a recurrent neural network (RNN). A technique of artificial intelligence, neural networks loosely mimic neurons of the mind to solve complex problems. He developed three RNN models for the prediction of roughness, rut, and alligator crack for asphalt pavements — performance metrics that he found to be sensitive to traffic weights in his paper.
To create his neural network, Guo developed a matrix of input layers that supply relevant data (such as pavement structures and freeze index), hidden layers that process and relate that data, and output layers that present the final calculations. Different from conventional feed-forward neural networks in pavement engineering, the hidden layers in RNN models can store key historical information for pavement deterioration.
Once Guo developed these models, he inputted road quality data from the FWHA’s Long Term Pavement Performance (LTTP) database. What he found was a clear relationship between traffic weight and certain forms of damage.
“My models show that increased traffic weights on asphalt pavements accelerate deterioration rates for roughness by 1.3 percent, rut by 7 percent, and alligator crack by 3.7 percent, given a representative asphalt pavement,” Guo explains.
Since he had a limited dataset, Guo’s model could not determine the role of traffic weight on other forms of damage for asphalt pavements. In the future, he will use more robust datasets to understand these other potential repercussions. He also hopes to explore the nationwide economic influence caused by overweight vehicles.  
Up until now, the specific impacts of traffic weight on road quality have been a loaded issue in the transportation community. Though some questions remain, Guo’s models have helped clarify a pernicious problem and helped advance an avenue for further, fruitful research.
The research was supported through the MIT Concrete Sustainability Hub by the Portland Cement Association and the Ready Mixed Concrete Research and Education Foundation. More

Contemporary robots can move quickly. “The motors are fast, and they’re powerful,” says Sabrina Neuman.
Yet in complex situations, like interactions with people, robots often don’t move quickly. “The hang up is what’s going on in the robot’s head,” she adds.
Perceiving stimuli and calculating a response takes a “boatload of computation,” which limits reaction time, says Neuman, who recently graduated with a PhD from the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL). Neuman has found a way to fight this mismatch between a robot’s “mind” and body. The method, called robomorphic computing, uses a robot’s physical layout and intended applications to generate a customized computer chip that minimizes the robot’s response time.
The advance could fuel a variety of robotics applications, including, potentially, frontline medical care of contagious patients. “It would be fantastic if we could have robots that could help reduce risk for patients and hospital workers,” says Neuman.
Neuman will present the research at this April’s International Conference on Architectural Support for Programming Languages and Operating Systems. MIT co-authors include graduate student Thomas Bourgeat and Srini Devadas, the Edwin Sibley Webster Professor of Electrical Engineering and Neuman’s PhD advisor. Other co-authors include Brian Plancher, Thierry Tambe, and Vijay Janapa Reddi, all of Harvard University. Neuman is now a postdoctoral NSF Computing Innovation Fellow at Harvard’s School of Engineering and Applied Sciences.
There are three main steps in a robot’s operation, according to Neuman. The first is perception, which includes gathering data using sensors or cameras. The second is mapping and localization: “Based on what they’ve seen, they have to construct a map of the world around them and then localize themselves within that map,” says Neuman. The third step is motion planning and control — in other words, plotting a course of action.
These steps can take time and an awful lot of computing power. “For robots to be deployed into the field and safely operate in dynamic environments around humans, they need to be able to think and react very quickly,” says Plancher. “Current algorithms cannot be run on current CPU hardware fast enough.”
Neuman adds that researchers have been investigating better algorithms, but she thinks software improvements alone aren’t the answer. “What’s relatively new is the idea that you might also explore better hardware.” That means moving beyond a standard-issue CPU processing chip that comprises a robot’s brain — with the help of hardware acceleration.
Hardware acceleration refers to the use of a specialized hardware unit to perform certain computing tasks more efficiently. A commonly used hardware accelerator is the graphics processing unit (GPU), a chip specialized for parallel processing. These devices are handy for graphics because their parallel structure allows them to simultaneously process thousands of pixels. “A GPU is not the best at everything, but it’s the best at what it’s built for,” says Neuman. “You get higher performance for a particular application.” Most robots are designed with an intended set of applications and could therefore benefit from hardware acceleration. That’s why Neuman’s team developed robomorphic computing.
The system creates a customized hardware design to best serve a particular robot’s computing needs. The user inputs the parameters of a robot, like its limb layout and how its various joints can move. Neuman’s system translates these physical properties into mathematical matrices. These matrices are “sparse,” meaning they contain many zero values that roughly correspond to movements that are impossible given a robot’s particular anatomy. (Similarly, your arm’s movements are limited because it can only bend at certain joints — it’s not an infinitely pliable spaghetti noodle.)
The system then designs a hardware architecture specialized to run calculations only on the non-zero values in the matrices. The resulting chip design is therefore tailored to maximize efficiency for the robot’s computing needs. And that customization paid off in testing.
Hardware architecture designed using this method for a particular application outperformed off-the-shelf CPU and GPU units. While Neuman’s team didn’t fabricate a specialized chip from scratch, they programmed a customizable field-programmable gate array (FPGA) chip according to their system’s suggestions. Despite operating at a slower clock rate, that chip performed eight times faster than the CPU and 86 times faster than the GPU.
“I was thrilled with those results,” says Neuman. “Even though we were hamstrung by the lower clock speed, we made up for it by just being more efficient.”
Plancher sees widespread potential for robomorphic computing. “Ideally we can eventually fabricate a custom motion-planning chip for every robot, allowing them to quickly compute safe and efficient motions,” he says. “I wouldn’t be surprised if 20 years from now every robot had a handful of custom computer chips powering it, and this could be one of them.” Neuman adds that robomorphic computing might allow robots to relieve humans of risk in a range of settings, such as caring for covid-19 patients or manipulating heavy objects.
“This work is exciting because it shows how specialized circuit designs can be used to accelerate a core component of robot control,” says Robin Deits, a robotics engineer at Boston Dynamics who was not involved in the research. “Software performance is crucial for robotics because the real world never waits around for the robot to finish thinking.” He adds that Neuman’s advance could enable robots to think faster, “unlocking exciting behaviors that previously would be too computationally difficult.”
Neuman next plans to automate the entire system of robomorphic computing. Users will simply drag and drop their robot’s parameters, and “out the other end comes the hardware description. I think that’s the thing that’ll push it over the edge and make it really useful.”
This research was funded by the National Science Foundation, the Computing Research Agency, the CIFellows Project, and the Defense Advanced Research Projects Agency. More

Three MIT computer science faculty members have been elected as fellows of the Association for Computing Machinery (ACM).
The new fellows are among 95 ACM members recognized as the top 1 percent for their outstanding accomplishments in computing and information technology and/or outstanding service to ACM and the larger computing community. Fellows are nominated by their peers, with nominations reviewed by a distinguished selection committee.
Anantha Chandrakasan is dean of the School of Engineering and the Vannevar Bush Professor of Electrical Engineering and Computer Science. He leads the MIT Energy-Efficient Circuits and Systems Group, which works on a variety of projects such as ultra-low-power internet-of-things devices, energy-efficient processors, machine learning processors, hardware security for computing devices, and wireless systems. He was recognized as a 2020 ACM fellow for energy-efficient design methodologies and circuits that enable ultra-low-power wireless sensors and computing devices.
Alan Edelman is an applied mathematics professor for the Department of Mathematics, the Applied Computing Group leader for the Computer Science and Artificial Intelligence Laboratory, and co-founder of the Julia programming language. His research includes high-performance computing, numerical computation, linear algebra, random matrix theory, and scientific machine learning. He was recognized as a 2020 ACM fellow for contributions to algorithms and languages for numerical and scientific computing.
Samuel Madden is the MIT Schwarzman College of Computing Distinguished Professor of Computing. Madden’s research is in the area of database systems, focusing on database analytics and query processing, ranging from clouds to sensors to modern high-performance server architectures. He co-directs the Data Systems for AI Lab initiative and the Data Systems Group, investigating issues related to systems and algorithms for data focusing on applying new methodologies for processing data, including applying machine learning methods to data systems and engineering data systems for applying machine learning at scale. He was recognized as a 2020 ACM fellow for contributions to data management and sensor computing systems. More

The field of artificial intelligence is moving at a staggering clip, with breakthroughs emerging in labs across MIT. Through the Undergraduate Research Opportunities Program (UROP), undergraduates get to join in. In two years, the MIT Quest for Intelligence has placed 329 students in research projects aimed at pushing the frontiers of computing and artificial intelligence, and using these tools to revolutionize how we study the brain, diagnose and treat disease, and search for new materials with mind-boggling properties.
Rafael Gomez-Bombarelli, an assistant professor in the MIT Department of Materials Science and Engineering, has enlisted several Quest-funded undergraduates in his mission to discover new molecules and materials with the help of AI. “They bring a blue-sky open mind and a lot of energy,” he says. “Through the Quest, we had the chance to connect with students from other majors who probably wouldn’t have thought to reach out.”
Some students stay in a lab for just one semester. Others never leave. Nick Bonaker is now in his third year working with Tamara Broderick, an associate professor in the Department of Electrical Engineering and Computer Science, to develop assistive technology tools for people with severe motor impairments.
“Nick has continually impressed me and our collaborators by picking up tools and ideas so quickly,” she says. “I particularly appreciate his focus on engaging so carefully and thoughtfully with the needs of the motor-impaired community. He has very carefully incorporated feedback from motor-impaired users, our charity collaborators, and other academics.”
This fall, MIT Quest celebrated two years of sponsoring UROP students. We highlight four of our favorite projects from last semester below.
Squeezing more energy from the sun
The price of solar energy is dropping as technology for converting sunlight into energy steadily improves. Solar cells are now close to hitting 50 percent efficiency in lab experiments, but there’s no reason to stop there, says Sean Mann, a sophomore majoring in computer science.
In a UROP project with Giuseppe Romano, a researcher at MIT’s Institute for Soldier Nanotechnologies, Mann is developing a solar cell simulator that would allow deep learning algorithms to systematically find better solar cell designs. Efficiency gains in the past have been made by evaluating new materials and geometries with hundreds of variables. “Traditional ways of exploring new designs is expensive, because simulations only measure the efficiency of that one design,” says Mann. “It doesn’t tell you how to improve it, which means you need either expert knowledge or lots more experiments to improve on it.”
The goal of Mann’s project is to develop a so-called differentiable solar cell simulator that computes the efficiency of a cell and describes how tweaking certain parameters will improve efficiency. Armed with this information, AI can predict which adjustments from among a dizzying array of combinations will boost cell performance the most. “Coupling this simulator with a neural network designed to maximize cell efficiency will eventually lead to some really good designs,” he says.
Mann is currently building an interface between AI models and traditional simulators. The biggest challenge so far, he says, has been debugging the simulator, which solves differential equations. He pulled several all-nighters double-checking his equations and code until he found the bug: an array of numbers off by one, skewing his results. With that obstacle down, Mann is now looking for algorithms to help the solver converge more quickly, a crucial step toward efficient optimization.
Teaching neural networks physics to identify stress fractures
Sensors deep within the modern jet engine sound an alarm when something goes wrong. But diagnosing the precise failure is often impossible without tinkering with the engine itself. To get a clearer picture faster, engineers are experimenting with physics-informed deep learning algorithms to translate these sensor distress signals.
“It would be way easier to find the part that has something wrong with it, rather than take the whole engine apart,” says Julia Gaubatz, a senior majoring in aerospace engineering. “It could really save people time and money in industry.”
Gaubatz spent the fall programming physical constraints into a deep learning model in a UROP project with Raul Radovitzky, a professor in MIT’s Department of Aeronautics and Astronautics, graduate student Grégoire Chomette, and third-year student Parker Mayhew. Their goal is to analyze the high-frequency signals coming from, say, a jet engine shaft, to pinpoint where a part may be stressed and about to crack. They hope to identify the particular points of failure by training neural networks on numerical simulations of how materials break to understand the underlying physics.
Working from her off-campus apartment in Cambridge, Massachusetts, Gaubatz built a smaller, simplified version of their physics-informed model to make sure their assumptions were correct. “It’s easier to look at the weights the neural network is coming up with to understand its predictions,” she says. “It’s like a test to check that the model is doing what it should according to theory.”
She picked the project to try applying what she had learned in a course on machine learning to solid mechanics, which focuses on how materials deform and break under force. Engineers are just starting to incorporate deep learning into the field, she says, and “it’s exciting to see how a new mathematical concept may change how we do things.”
Training an AI to reason its way through visual problems
An artificial intelligence model that can play chess at superhuman levels may be hopeless at Sudoku. Humans, by contrast, pick up new games easily by adapting old knowledge to new environments. To give AI more of this flexibility, researchers created the ARC visual-reasoning dataset to motivate the field to create new techniques for solving problems involving abstraction and reasoning.
“If an AI does well on the test, it signals a more human-like intelligence,” says first-year student Subhash Kantamneni, who joined a UROP project this fall in the lab of Department of Brain and Cognitive Sciences (BSC) Professor Tomaso Poggio, which is part of the Center for Minds, Brains and Machines.
Poggio’s lab hopes to crack the ARC challenge by merging deep learning and automated program-writing to train an agent to solve ARC’s 400 tasks by writing its own programs. Much of their work takes place in DreamCoder, a tool developed at MIT that learns new concepts while solving specialized tasks. Using DreamCoder, the lab has so far solved 70 ARC tasks, and Kantamneni this fall worked with master of engineering student Simon Alford to tackle the rest.
To try and solve ARC’s 20 or so pattern-completion tasks, Kantamneni created a script to generate similar examples to train the deep learning model. He also wrote several mini programs, or primitives, to solve a separate class of tasks that involve performing logical operations on pixels. With the help of these new primitives, he says, DreamCoder learned to combine the old and new programs to solve ARC’s 10 or so pixelwise tasks.
The coding and debugging was hard work, he says, but the other lab members made him feel at home and appreciated. “I don’t think they even knew I was a freshman,” he says. “They listened to what I had to say and valued my input.”
Putting language comprehension under a microscope
Language is more than a system of symbols: It allows us to express concepts and ideas, think and reason, and communicate and coordinate with others. To understand how the brain does it, psychologists have developed methods for tracking how quickly people grasp what they read and hear. Longer reading times can indicate when a word has been improperly used, offering insight into how the brain incrementally finds meaning in a string of words.
In a UROP project this fall in Roger Levy’s lab in BCS, sophomore Pranali Vani ran a set of sentence-processing experiments online that were developed by an earlier UROP student. In each sentence, one word is placed in such a way that it creates an impression of ambiguity or implausibility. The weirder the sentence, the longer it takes a human subject to decipher its meaning. For example, placing a verb like “tripped” at the end of a sentence, as in “The woman brought the sandwich from the kitchen tripped,” tends to throw off native English speakers. Though grammatically correct, the wording implies that bringing rather than tripping is the main action of the sentence, creating confusion for the reader.
In three sets of experiments, Vani found that the biggest slowdowns came when the verb was positioned in a way that sounded ungrammatical. Vani and her advisor, Ethan Wilcox, a PhD student at Harvard University, got similar results when they ran the experiments on a deep learning model.
“The model was ‘surprised’ when the grammatical interpretation is unlikely,” says Wilcox. Though the model isn’t explicitly trained on English grammar, he says, the results suggest that a neural network trained on reams of text effectively learns the rules anyway.
Vani says she enjoyed learning how to program in R and shell scripts like Dash. She also gained an appreciation for the persistence needed to conduct original research. “It takes a long time,” she says. “There’s a lot of thought that goes into each detail and each decision made during the course of an experiment.”
Funding for MIT Quest UROP projects this fall was provided, in part, by the MIT-IBM Watson AI Lab. More