Written by Adam Zewe, MIT News Office
A new academic program developed at MIT aims to teach U.S. Air and Space Forces personnel to understand and utilize artificial intelligence technologies. In a recent peer-reviewed study, the program researchers found that this approach was effective and well-received by employees with diverse backgrounds and professional roles.
The project, which was funded by the Department of the Air Force–MIT Artificial Intelligence Accelerator, seeks to contribute to AI educational research, specifically regarding ways to maximize learning outcomes at scale for people from a variety of educational backgrounds.
Experts in MIT Open Learning built a curriculum for three general types of military personnel — leaders, developers, and users — utilizing existing MIT educational materials and resources. They also created new, more experimental courses that were targeted at Air and Space Forces leaders.
Then, MIT scientists led a research study to analyze the content, evaluate the experiences and outcomes of individual learners during the 18-month pilot, and propose innovations and insights that would enable the program to eventually scale up.
They used interviews and several questionnaires, offered to both program learners and staff, to evaluate how 230 Air and Space Forces personnel interacted with the course material. They also collaborated with MIT faculty to conduct a content gap analysis and identify how the curriculum could be further improved to address the desired skills, knowledge, and mindsets.
Ultimately, the researchers found that the military personnel responded positively to hands-on learning; appreciated asynchronous, time-efficient learning experiences to fit in their busy schedules; and strongly valued a team-based, learning-through-making experience but sought content that included more professional and soft skills. Learners also wanted to see how AI directly applied to their day-to-day work and the broader mission of the Air and Space Forces. They were also interested in more opportunities to engage with others, including their peers, instructors, and AI experts.
Based on these findings, which the program researchers recently shared at the IEEE Frontiers in Education Conference, the team is augmenting the educational content and adding new technical features to the portal for the next iteration of the study, which is currently underway and will extend through 2023.
“We are digging deeper into expanding what we think the opportunities for learning are, that are driven by our research questions but also from understanding the science of learning about this kind of scale and complexity of a project. But ultimately we are also trying to deliver some real translational value to the Air Force and the Department of Defense. This work is leading to a real-world impact for them, and that is really exciting,” says principal investigator Cynthia Breazeal, who is MIT’s dean for digital learning, director of MIT RAISE (Responsible AI for Social Empowerment and Education), and head of the Media Lab’s Personal Robots research group.
Building learning journeys
At the outset of the project, the Air Force gave the program team a set of profiles that captured educational backgrounds and job functions of six basic categories of Air Force personnel. The team then created three archetypes it used to build “learning journeys” — a series of training programs designed to impart a set of AI skills for each profile.
The Lead-Drive archetype is an individual who is making strategic decisions; the Create-Embed archetype is a technical worker who is implementing AI solutions; and the Facilitate-Employ archetype is an end-user of AI-augmented tools.
It was a priority to convince the Lead-Drive archetype of the importance of this program, says lead author Andrés Felipe Salazar-Gomez, a research scientist at MIT Open Learning.
“Even inside the Department of Defense, leaders were questioning if training in AI is worth it or not,” he explains. “We first needed to change the mindset of the leaders so they would allow the other learners, developers, and users to go through this training. At the end of the pilot we found they embraced this training. They had a different mindset.”
The three learning journeys, which ranged from six to 12 months, included a combination of existing AI courses and materials from MIT Horizon, MIT Lincoln Laboratory, MIT Sloan School of Management, the Computer Science and Artificial Intelligence Laboratory (CSAIL), the Media Lab, and MITx MicroMasters programs. Most educational modules were offered entirely online, either synchronously or asynchronously.
Each learning journey included different content and formats based on the needs of users. For instance, the Create-Embed journey included a five-day, in-person, hands-on course taught by a Lincoln Laboratory research scientist that offered a deep dive into technical AI material, while the Facilitate-Employ journey comprised self-paced, asynchronous learning experiences, primarily drawing on MIT Horizon materials that are designed for a more general audience.
The researchers also created two new courses for the Lead-Drive cohort. One, a synchronous online course called The Future of Leadership: Human and AI Collaboration in the Workforce, developed in collaboration with Esme Learning, was based on the leaders’ desire for more training around ethics and human-centered AI design and more content on human-AI collaboration in the workforce. The researchers also crafted an experimental, three-day, in-person course called Learning Machines: Computation, Ethics, and Policy that immersed leaders in a constructionist-style learning experience where teams worked together on a series of hands-on activities with autonomous robots that culminated in an escape-room style capstone competition that brought everything together.
The Learning Machines course was wildly successful, Breazeal says.
“At MIT, we learn by making and through teamwork. We thought, what if we let executives learn about AI this way?” she explains. “We found that the engagement is much deeper, and they gained stronger intuitions about what makes these technologies work and what it takes to implement them responsibly and robustly. I think this is going to deeply inform how we think about executive education for these kinds of disruptive technologies in the future.”
Gathering feedback, enhancing content
Throughout the study, the MIT researchers checked in with the learners using questionnaires to obtain their feedback on the content, pedagogies, and technologies used. They also had MIT faculty analyze each learning journey to identify educational gaps.
Overall, the researchers found that the learners wanted more opportunities to engage, either with their peers through team-based activities or with faculty and experts through synchronous components of online courses. And while most personnel found the content to be interesting, they wanted to see more examples that were directly applicable to their day-to-day work.
Now in the second iteration of the study, researchers are using that feedback to enhance the learning journeys. They are designing knowledge checks that will be a part of the self-paced, asynchronous courses to help learners engage with the content. They are also adding new tools to support live Q&A events with AI experts and help build more community among learners.
The team is also looking to add specific Department of Defense examples throughout the educational modules, and include a scenario-based workshop.
“How do you upskill a workforce of 680,000 across diverse work roles, all echelons, and at scale? This is an MIT-sized problem, and we are tapping into the world-class work that MIT Open Learning has been doing since 2013 — democratizing education on a global scale,” says Maj. John Radovan, deputy director of the DAF-MIT AI Accelerator. “By leveraging our research partnership with MIT, we are able to research the optimal pedagogy of our workforce through focused pilots. We are then able to quickly double down on unexpected positive results and pivot on lessons learned. This is how you accelerate positive change for our airmen and guardians.”
As the study progresses, the program team is sharpening their focus on how they can enable this training program to reach a larger scale.
“The U.S. Department of Defense is the largest employer in the world. When it comes to AI, it is really important that their employees are all speaking the same language,” says Kathleen Kennedy, senior director of MIT Horizon and executive director of the MIT Center for Collective Intelligence. “But the challenge now is scaling this so that learners who are individual people get what they need and stay engaged. And this will certainly help inform how different MIT platforms can be used with other types of large groups.”
Study: Superconductivity switches on and off in “magic-angle” graphene
A quick electric pulse completely flips the material’s electronic properties, opening a route to ultrafast, brain-inspired, superconducting electronics
Written by Jennifer Chu, MIT News Office
With some careful twisting and stacking, MIT physicists have revealed a new and exotic property in “magic-angle” graphene: superconductivity that can be turned on and off with an electric pulse, much like a light switch.
The discovery could lead to ultrafast, energy-efficient superconducting transistors for neuromorphic devices — electronics designed to operate in a way similar to the rapid on/off firing of neurons in the human brain.
Magic-angle graphene refers to a very particular stacking of graphene — an atom-thin material made from carbon atoms that are linked in a hexagonal pattern resembling chicken wire. When one sheet of graphene is stacked atop a second sheet at a precise “magic” angle, the twisted structure creates a slightly offset “moiré” pattern, or superlattice, that is able to support a host of surprising electronic behaviors.
In 2018, Pablo Jarillo-Herrero and his group at MIT were the first to demonstrate magic-angle twisted bilayer graphene. They showed that the new bilayer structure could behave as an insulator, much like wood, when they applied a certain continuous electric field. When they upped the field, the insulator suddenly morphed into a superconductor, allowing electrons to flow, friction-free.
That discovery gave rise to “twistronics,” a field that explores how certain electronic properties emerge from the twisting and layering of two-dimensional materials. Researchers including Jarillo-Herrero have continued to reveal surprising properties in magic-angle graphene, including various ways to switch the material between different electronic states. So far, such “switches” have acted more like dimmers, in that researchers must continuously apply an electric or magnetic field to turn on superconductivity, and keep it on.
Now Jarillo-Herrero and his team have shown that superconductivity in magic-angle graphene can be switched on, and kept on, with just a short pulse rather than a continuous electric field. The key, they found was a combination of twisting and stacking.
In a paper appearing today in Nature Nanotechnology, the team reports that, by stacking magic-angle graphene between two offset layers of boron nitride — a two-dimensional insulating material — the unique alignment of the sandwich structure enabled the researchers to turn graphene’s superconductivity on and off with a short electric pulse.
“For the vast majority of materials, if you remove the electric field, zzzzip, the electric state is gone,” says Jarillo-Herrero, who is the Cecil and Ida Green Professor of Physics at MIT. “This is the first time that a superconducting material has been made that can be electrically switched on and off, abruptly. This could pave the way for a new generation of twisted, graphene-based superconducting electronics.”
His MIT co-authors are lead author Dahlia Klein, Li-Qiao Xia, and David MacNeill, along with Kenji Watanabe and Takashi Taniguchi of the National Institute for Materials Science in Japan.
Flipping the switch
In 2019, a team at Stanford University discovered that magic-angle graphene could be coerced into a ferromagnetic state. Ferromagnets are materials that retain their magnetic properties, even in the absence of an externally applied magnetic field.
The researchers found that magic-angle graphene could exhibit ferromagnetic properties in a way that could be tuned on and off. This happened when the graphene sheets were layered between two sheets of boron nitride such that the crystal structure of the graphene was aligned to one of the boron nitride layers. The arrangement resembled a cheese sandwich in which the top slice of bread and the cheese orientations are aligned, but the bottom slice of bread is rotated at a random angle with respect to the top slice. The result intrigued the MIT group.
“We were trying to get a stronger magnet by aligning both slices,” Jarillo-Herrero says. “Instead, we found something completely different.”
In their current study, the team fabricated a sandwich of carefully angled and stacked materials. The “cheese” of the sandwich consisted of magic-angle graphene — two graphene sheets, the top rotated slightly at the “magic” angle of 1.1 degrees with respect to the bottom sheet. Above this structure, they placed a layer of boron nitride, exactly aligned with the top graphene sheet. Finally, they placed a second layer of boron nitride below the entire structure and offset it by 30 degrees with respect to the top layer of boron nitride.
The team then measured the electrical resistance of the graphene layers as they applied a gate voltage. They found, as others have, that the twisted bilayer graphene switched electronic states, changing between insulating, conducting, and superconducting states at certain known voltages.
What the group did not expect was that each electronic state persisted rather than immediately disappearing once the voltage was removed — a property known as bistability. They found that, at a particular voltage, the graphene layers turned into a superconductor, and remained superconducting, even as the researchers removed this voltage.
This bistable effect suggests that superconductivity can be turned on and off with short electric pulses rather than a continuous electric field, similar to flicking a light switch. It isn’t clear what enables this switchable superconductivity, though the researchers suspect it has something to do with the special alignment of the twisted graphene to both boron nitride layers, which enables a ferroelectric-like response of the system. (Ferroelectric materials display bistability in their electric properties.)
“By paying attention to the stacking, you could add another tuning knob to the growing complexity of magic-angle, superconducting devices,” Klein says.
For now, the team sees the new superconducting switch as another tool researchers can consider as they develop materials for faster, smaller, more energy-efficient electronics.
“People are trying to build electronic devices that do calculations in a way that’s inspired by the brain,” Jarillo-Herrero says. “In the brain, we have neurons that, beyond a certain threshold, they fire. Similarly, we now have found a way for magic-angle graphene to switch superconductivity abruptly, beyond a certain threshold. This is a key property in realizing neuromorphic computing.”
This research was supported in part by the Air Force Office of Scientific Research, the Army Research Office, and the Gordon and Betty Moore Foundation.
When should data scientists try a new technique?
A new measure can help scientists decide which estimation method to use when modeling a particular data problem
Written by Adam Zewe, MIT News Office
If a scientist wanted to forecast ocean currents to understand how pollution travels after an oil spill, she could use a common approach that looks at currents traveling between 10 and 200 kilometers. Or, she could choose a newer model that also includes shorter currents. This might be more accurate, but it could also require learning new software or running new computational experiments. How to know if it will be worth the time, cost, and effort to use the new method?
A new approach developed by MIT researchers could help data scientists answer this question, whether they are looking at statistics on ocean currents, violent crime, children’s reading ability, or any number of other types of datasets.
The team created a new measure, known as the “c-value,” that helps users choose between techniques based on the chance that a new method is more accurate for a specific dataset. This measure answers the question “is it likely that the new method is more accurate for this data than the common approach?”
Traditionally, statisticians compare methods by averaging a method’s accuracy across all possible datasets. But just because a new method is better for all datasets on average doesn’t mean it will actually provide a better estimate using one particular dataset. Averages are not application-specific.
So, researchers from MIT and elsewhere created the c-value, which is a dataset-specific tool. A high c-value means it is unlikely a new method will be less accurate than the original method on a specific data problem.
In their proof-of-concept paper, the researchers describe and evaluate the c-value using real-world data analysis problems: modeling ocean currents, estimating violent crime in neighborhoods, and approximating student reading ability at schools. They show how the c-value could help statisticians and data analysts achieve more accurate results by indicating when to use alternative estimation methods they otherwise might have ignored.
“What we are trying to do with this particular work is come up with something that is data specific. The classical notion of risk is really natural for someone developing a new method. That person wants their method to work well for all of their users on average. But a user of a method wants something that will work on their individual problem. We’ve shown that the c-value is a very practical proof-of-concept in that direction,” says senior author Tamara Broderick, an associate professor in the Department of Electrical Engineering and Computer Science (EECS) and a member of the Laboratory for Information and Decision Systems and the Institute for Data, Systems, and Society.
She’s joined on the paper by Brian Trippe PhD ’22, a former graduate student in Broderick’s group who is now a postdoc at Columbia University; and Sameer Deshpande ’13, a former postdoc in Broderick’s group who is now an assistant professor at the University of Wisconsin at Madison. An accepted version of the paper is posted online in the Journal of the American Statistical Association.
The c-value is designed to help with data problems in which researchers seek to estimate an unknown parameter using a dataset, such as estimating average student reading ability from a dataset of assessment results and student survey responses. A researcher has two estimation methods and must decide which to use for this particular problem.
The better estimation method is the one that results in less “loss,” which means the estimate will be closer to the ground truth. Conder again the forecasting of ocean currents: Perhaps being off by a few meters per hour isn’t so bad, but being off by many kilometers per hour makes the estimate useless. The ground truth is unknown, though; the scientist is trying to estimate it. Therefore, one can never actually compute the loss of an estimate for their specific data. That’s what makes comparing estimates challenging. The c-value helps a scientist navigate this challenge.
The c-value equation uses a specific dataset to compute the estimate with each method, and then once more to compute the c-value between the methods. If the c-value is large, it is unlikely that the alternative method is going to be worse and yield less accurate estimates than the original method.
“In our case, we are assuming that you conservatively want to stay with the default estimator, and you only want to go to the new estimator if you feel very confident about it. With a high c-value, it’s likely that the new estimate is more accurate. If you get a low c-value, you can’t say anything conclusive. You might have actually done better, but you just don’t know,” Broderick explains.
Probing the theory
The researchers put that theory to the test by evaluating three real-world data analysis problems.
For one, they used the c-value to help determine which approach is best for modeling ocean currents, a problem Trippe has been tackling. Accurate models are important for predicting the dispersion of contaminants, like pollution from an oil spill. The team found that estimating ocean currents using multiple scales, one larger and one smaller, likely yields higher accuracy than using only larger scale measurements.
“Oceans researchers are studying this, and the c-value can provide some statistical ‘oomph’ to support modeling the smaller scale,” Broderick says.
In another example, the researchers sought to predict violent crime in census tracts in Philadelphia, an application Deshpande has been studying. Using the c-value, they found that one could get better estimates about violent crime rates by incorporating information about census-tract-level nonviolent crime into the analysis. They also used the c-value to show that additionally leveraging violent crime data from neighboring census tracts in the analysis isn’t likely to provide further accuracy improvements.
“That doesn’t mean there isn’t an improvement, that just means that we don’t feel confident saying that you will get it,” she says.
Now that they have proven the c-value in theory and shown how it could be used to tackle real-world data problems, the researchers want to expand the measure to more types of data and a wider set of model classes.
The ultimate goal is to create a measure that is general enough for many more data analysis problems, and while there is still a lot of work to do to realize that objective, Broderick says this is an important and exciting first step in the right direction.
This research was supported, in part, by an Advanced Research Projects Agency-Energy grant, a National Science Foundation CAREER Award, the Office of Naval Research, and the Wisconsin Alumni Research Foundation.
How to push, wiggle, or drill an object through sand
A method for quickly predicting the forces needed to push objects through soft, granular materials could help engineers drive robots or anchor ships
Written by Jennifer Chu, MIT News Office
Pushing a shovel through snow, planting an umbrella on the beach, wading through a ball pit, and driving over gravel all have one thing in common: They all are exercises in intrusion, with an intruding object exerting some force to move through a soft and granular material.
Predicting what it takes to push through sand, gravel, or other soft media can help engineers drive a rover over Martian soil, anchor a ship in rough seas, and walk a robot through sand and mud. But modeling the forces involved in such processes is a huge computational challenge that often takes days to weeks to solve.
Now, engineers at MIT and Georgia Tech have found a faster and simpler way to model intrusion through any soft, flowable material. Their new method quickly maps the forces it would take to push, wiggle, and drill an object through granular material in real-time. The method can apply to objects and grains of any size and shape, and does not require complex computational tools as other methods do.
“We now have a formula that can be very useful in settings where you have to check through lots of options as fast as possible,” says Ken Kamrin, professor of mechanical engineering at MIT.
“This is especially useful for applications such as real-time path-planning for vehicles traveling through vast deserts and other off-road terrains, that cannot wait for existing slower simulation methods to decide their path,” adds Shashank Agarwal SM ’19, PhD ’22.
Kamrin and Agarwal detail their new method in a study appearing this week in the journal Proceedings of the National Academy of Sciences. The study also includes Daniel I. Goldman, professor of physics at Georgia Tech.
A fluid connection
In order to know how much to push on an object to move it through sand, one could go grain by grain, using discrete element modeling, or DEM — an approach that systematically calculates each individual grain’s motion in response to a given force. DEM is precise but slow, and it can take weeks to fully solve a practical problem involving just a handful of sand. As a faster alternative, scientists can develop continuum models, which simulate granular behavior in generalized chunks, or grain groupings. This more simplified approach can still generate a detailed picture of how grains flow, in a way that can shave a weeks-long problem down to days or even hours.
“We wanted to see if we could do even better than that and cut that process down to seconds,” Agarwal says.
The team looked to previous work by Goldman. In 2014, he was studying how animals and robots move through dry, granular material such as sand and soil. In looking for ways to quantitatively describe their movements, he found he could do so with a quick relationship that was originally meant to describe fluid swimmers.
The formulation, Resistive Force Theory (RFT), works by considering an object’s surface as a collection of small plates. (Imagine representing a sphere as a soccer ball.) As an object moves through a fluid, each plate experiences a force, and RFT claims that the force on each plate depends only on its local orientation and movement. The equation takes all this into account, along with the fluid’s individual characteristics, to ultimately describe how the object as a whole moves through a fluid.
Surprisingly, Goldman found this simple approach was also accurate when applied to granular intrusion. Specifically, it predicted the forces lizards and snakes exert to slither through sand, as well as how small, legged robots walk over soil. The question, Kamrin says, was why?
“It was this weird mystery why this theory, which was originally derived for moving through viscous fluid, would even work at all in granular media, which has completely different flow behavior,” he says.
Kamrin took a closer look at the math and found a connection between RFT and a continuum model he had derived to describe granular flow. In other words, the physics checked out, and RFT could indeed be an accurate way to predict granular flow, in a simpler and faster way than conventional models. But there was one big limitation: The approach was mainly workable for two-dimensional problems.
To model intrusion using RFT, one needs to know what will happen if one moves a plate every which way possible — a task that is manageable in two dimensions, but not in three. The team then needed some shortcut to simplify 3D’s complexity.
In their new study, the researchers adapted RFT to 3D by adding an extra ingredient to the equation. That ingredient is a plate’s twist angle, measuring how plate orientation changes as the entire object is rotated. When they incorporated this extra angle, in addition to a plate’s tilt and direction of motion, the team had enough information to define the force acting on the plate as it moves through a material in 3D. Importantly, by exploiting the connection to continuum modeling, the resulting 3D-RFT is generalizable, and can be easily recalibrated to apply to many dry granular media on Earth, and even on other planetary bodies.
The researchers demonstrated the new method using a variety of three-dimensional objects, from simple cylinders and cubes to more complex bunny- and monkey-shaped geometries. They first tiled the objects, representing them each as a collection of hundreds to thousands of tiny plates. Then they applied the tweaked RFT formula to each individual plate and calculated the forces that would be needed over time to drill each plate, and ultimately the entire object, down through a bed of sand.
“For more wacky objects, like the bunny, you can imagine having to consistently shift your loads to keep drilling it straight down,” Kamrin says. “And our method can even predict those little wiggles, and the distribution of force all around the bunny, in less than a minute.”
The new approach provides a fast and accurate way to model granular intrusion, which can be applied to a host of practical problems, from driving a rover through Martian soil, to characterizing the movement of animals through sand, and even predicting what it would take to uproot a tree.
“Can I predict how hard it is to uproot natural plants? You might want to know, is this storm going to knock over this tree?” Kamrin says. “Here is a way to get an answer fast.”
This research was supported, in part, by the Army Research Office, the U.S. Army DEVCOM Ground Vehicle Systems Center, and NASA.
Business & Economy9 months ago
NSE Academy Limited collaborates with HDFC Mutual Fund for financial awareness program
Edu News9 months ago
Technique protects privacy when making online recommendations
Business & Economy6 months ago
Using artificial intelligence to control digital manufacturing
Edu News8 months ago
Stronger security for smart devices
Edu News8 months ago
Astronomers discover a multiplanet system nearby
Edu News9 months ago
Search reveals eight new sources of black hole echoes
Edu News7 months ago
Jasudben ML School celebrated its first edition of Pride Month
Edu News9 months ago
Unpacking black-box models