Mapping the ocean with autonomous sensors

In late October 2025, Tropical Storm Melissa moved through the Caribbean Sea with moderate winds that didn’t get much attention. But on Oct. 25, aided by a patch of warm ocean, the storm rapidly intensified. By the time it made landfall in Jamaica, it was one of the strongest Atlantic hurricanes on record, uprooting trees, tearing the roofs from buildings, and causing catastrophic flooding and power outages.

Ravi Pappu SM ’95, PhD ’01 blames the surprise on our inability to gather high-quality ocean data.

“The storm intensified because of a small pool of hot water in the Caribbean Ocean that fed it energy,” Pappu explains. “These pools are everywhere. They can be hundreds of kilometers wide and are literally invisible to us. If we knew about that pool, we could say very precisely how the hurricane would intensify and better deal with it.”

Pappu thinks he has a way to solve that problem. He is the founder of Apeiron Labs, a company deploying low-cost autonomous ocean sensors to capture more data, in more places, and at a lower cost than is possible today. The company’s devices roam the ocean up to a quarter mile below the surface and continuously gather data on temperature, acoustics, salinity, and more, providing a real-time look at one of the planet’s last known mysteries. He says the sensors can do for the ocean what small, modular CubeSat satellites did for Earth observation from space.

When the devices are ready to be recharged, trackers make it easy to scoop them from the ocean surface. Pappu envisions the recovery process being done by autonomous boats in the future.

“Humanity needs ocean measurements, and we need them at a scale that has never been attempted before,” Pappu says. “It’s a massively hard problem. In the last century, oceanographers resigned themselves to calling it the century of undersampling. If we are successful, we will have a much more fine-grained understanding of our oceans and how they impact humans. That’s what drives us.”

Homework

Pappu came to MIT after completing a 10-year homework assignment. It started when he was a child in India in the 1980s, when he saw a hologram on the cover of National Geographic for the first time.

“I was so taken by it that I decided I needed to learn how to make those three-dimensional images,” Pappu recalls. “I learned what I could by reading books and papers. I didn’t know who invented the hologram until I read a book about MIT’s Media Lab. The book named the person who invented the rainbow hologram, so I wrote him a letter. I didn’t know his address, so I just wrote on the envelope, ‘Steve Benton, holography researcher, MIT, USA.’”

To Pappu’s surprise, the letter reached Benton, and the former Media Lab professor even wrote back with some further topics he needed to learn about.

Pappu never forgot that. He earned a bachelor’s degree in electrical engineering in India, then earned his master’s degree at Villanova University, taking all the optics classes he could.

“Eventually, about 10 years after I saw my first hologram, I wrote to Steve and I said, ‘I did all these things you asked me, now I want to study with you,’” Pappu says. “That’s how I got into MIT.”

Pappu studied under Benton for the next three years. He also studied under Professor Neil Gershenfeld as part of his PhD. Following graduation, Pappu and four classmates started ThingMagic, a consulting company that eventually sold RFID readers. ThingMagic was acquired 2010. Pappu returned to MIT for two years as a visiting scientist around the time of the acquisition.

Following that experience, Pappu worked at In-Q-Tel, an organization that invested in ThingMagic and other companies with potential to advance national security. It was there that Pappu realized how badly the world needed large-scale, inexpensive ocean sensing.

“All of the ocean sensing up to that point, and even today, was about making a really expensive thing that cost $20 million, goes to the bottom of the ocean, and stays there for five years,” Pappu says. “We needed things that are cheap and scalable to deploy wherever you need them for as long as you want.”

Pappu officially founded Apeiron Labs in 2022.

“What we’re focused on is figuring out how the ocean works,” Pappu says. “How warm is it? What is the pH? How salty is it? These things vary from place to place every 10 kilometers or so. It varies over time, and it varies by season. If we knew the details of the ocean with the same fidelity we have for the atmosphere, we would be able to tell exactly when and where hurricanes hit. It would mean less uncertainty.”

Apeiron’s ocean-sensing devices are each 3 feet long and about 20 pounds. They’re designed to be dropped off a boat or plane with biodegradable parachutes and stay in the ocean for six months. Each device continuously sends data to the cloud, is controllable through a cloud-based ocean operating system, and is accessible on a mobile phone.

“We lower the carbon footprint and cost of gathering ocean data because everything else needs a diesel ship — and a fully crewed ship costs $100,000 a day,” Rappu says. “By the time you collect the first data in the old model, you’ve already committed to a lot of money in addition to millions of dollars for the sensors. “

The company’s devices currently have two types of sensors: one for measuring salinity, temperature, and depth, and the other that uses a hydrophone to passively listen for things like submarines and whales.

That could be used to detect the low-frequency calls and clicks of endangered whales and other fish species. Currently, fishermen must look for whales manually with spotters on ships or planes. The data could also be used to improve weather forecasts, monitor noise from offshore energy projects, and track currents.

“Currents are determined by temperature and salinity, so if there’s an oil spill, our data could help determine where that spill is going,” Pappu says. “Or if you’re a fisherman, knowing where the water changes from warm to cold, which is where the fish hang out, is very useful.”

An ocean of possibilities

Apeiron Labs has worked with government defense agencies including the U.S. Navy over the last two years. The company has also tested its devices off the coast of California and in the Boston Harbor.

“The most important thing is, when we show people our approach and what we’ve demonstrated so far, they are no longer asking, ‘Can it be done?’ they’re asking, ‘What can we do with it?’” Pappu says. “Our customers have spent decades working in the ocean and they understand how novel these capabilities are.”

Of all the possibilities, improved storm forecasting could be the one Pappu is most excited about.

“Our mission is to lower the barriers to ocean data,” Pappu says. “The ocean is a huge determinant of weather, climate, and short-term forecasting. Despite our best efforts to predict the intensity of storms, sudden changes are still the norm, and much of that comes down to a lack of understanding of our oceans. If we were monitoring these things over long periods of time and finer spatial scales, we could see these storms coming much earlier with more certainty.”

de MIT News https://ift.tt/p1lxtk8

Rethinking how our brains use categories to make sense of the world

In the new review article, “Categorization is Baked into the Brain,” cognitive scientists Earl K. Miller, Picower Professor of Neuroscience at MIT, and Lisa Feldman Barrett, university distinguished professor at Northeastern University, contend that categorization is part of a predictive process the brain uses to efficiently meet the body’s needs in a fast-paced, otherwise overwhelming sensory world. In that sense, their paper in Nature Reviews Neuroscience challenges decades of dogma about how and why the brain boils down what it sees, hears, smells, tastes, and feels.

Categories are groups of things that are similar enough to be considered functionally equivalent. When you walk through a neighborhood, you’ll naturally experience the furry, four-legged, barking animal ahead of you as a “dog.” In the classic view of cognition, your brain arrives at that categorization by soaking in lots of basic sensory features of the hound — its shape, its size, the sounds it makes, its behavior — and compares that to some prototype “dog” stored in your memory. Hundreds of milliseconds after the first sensory inputs, you can then decide what you might want to do about the dog.

Barrett and Miller argue that that’s wrong. Instead, they propose that your brain comes prepared for sensory patterns with predictions of the motor action plans that are most likely to achieve the needs and goals you bring to the moment. Those prediction signals can be described as a momentary category that the brain constructs to shape the processing of sensory signals. 

From the very start, incoming sensory signals are compressed and abstracted into that category to efficiently select the best predicted plan. If you are in an unfamiliar neighborhood your brain might construct the category “dog” to avoid being bit, resulting in: “Back away slowly while saying nice doggie.” If you are on your own block and encounter a familiar dog, your brain might construct a category to kneel and open up your arms to summon your neighbor’s adorable pup for a satisfying petting.

In either case, the category “dog” arises in the context of your needs and your prediction from a menu of learned action plans for similar situations, not from an intellectual exercise of neutrally regarding sensory inputs, comparing them to a fixed prototype, and then planning from there. If the brain really worked the classically believed way, you’d be on the back foot when the unfamiliar dog lunged at you.

“One of the main things your brain has to do is predict the world,” says Miller, a faculty member of The Picower Institute for Learning and Memory and the Department of Brain and Cognitive Sciences at MIT. “It takes several hundred milliseconds to process things, and meanwhile the world is moving on. Your brain has to anticipate things.”

The most pragmatic and efficient way to survive and thrive in such a world, Barrett says, is to have your needs and potential plans ready for the sensory situation. If your predictions are right, you’re prepared in time. If they are wrong, you adjust and learn from it.

“The stimulus, cognition, response model of the brain is wrong,” says Barrett, a faculty member in Northeastern’s Department of Psychology and co-director of the Interdisciplinary Affective Science Laboratory. “The brain prepares for a response and then perceives a stimulus. A brain is not reactive. It’s predictive. Action planning comes first. Perception comes second, as a function of the action plan.”

Anatomical and functional evidence

Throughout the review, Barrett and Miller ground the provocative proposal in copious anatomical, electrophysiological, and imaging evidence about the brain. They cite numerous experimental results that show how the brain is structured to broadcast memories to create motor plans that flow back toward signals that arrive from the body’s sensory surfaces, actively whittling them down and shaping them to give them meaning.

“The capacity to create similarities from differences — to abstract — is embedded in the architecture of the nervous system, and you can see that by looking at what is connected to what and by observing signal flow,” Barrett says.

For example, as circuits feed signals “forward” from sensory surfaces (such as the retina) to regions of the cerebral cortex that are focused on sensory processing (such as the visual cortex) toward the areas that are important for executive control (the prefrontal cortex) and control of the body (limbic cortex), information passes from many small, barely connected neurons to fewer, bigger, and more well-connected neurons. Such an architecture compresses sensory details into increasingly abstract representations that group many different features into smaller groups of similar features, and in doing so helps to select a predicted action plan from the broader category that’s already there.

“Your brain is a big funnel to take the outside world and turn it into an output,” Miller says.

Moreover, anatomical evidence shows that the neurons in the cortex maintain many more connections to provide feedback from memory that control sensory regions than to feed sensory information forward. As much as 90 percent of synapses in the visual cortex are “feedback” instead of “feedforward,” Barrett and Miller wrote. In other words, the brain is built to use memory to filter incoming sensory signals, consistent with imposing needs and goals on what would otherwise be a deluge of sights, sounds, and other sensations.

Yet another line of evidence are numerous studies from Miller’s own lab showing that at the broad network level of information flow in the cortex, the brain uses beta frequency waves that carry information about goals and plans, to constrain the expression of gamma frequency waves that carry information about specific sensory inputs.

Finally, the dominance of “feedback” over “feedforward” signals in the cortical architecture allows for the possibility that sensory signals are made meaningful in terms of predicted plans. When these plans are wrong, the resulting surprise can be integrated for future use.

“In science, there is a special name for that: learning,” Barrett says.

Implications for human thought and disease

In the end, Barrett and Miller’s proposal completely changes the idea of categorization, shifting it from being a particular intellectual skill to being a fundamental function for predictively meeting the body’s needs (or, “allostasis”).

“A category may not be a representation that an animal has, but a signal processing event than an animal does, predictively, to constrain the meaning of a high-dimensional ensemble of signals in a particular situation,” the authors wrote. “Categorization renders these signals meaningful — similar to one another and to past allostatic events — in terms of some goal or function.”

Humans, Barrett says, have a relatively massive amount of the neural network architecture to perform these pragmatic abstractions, and therefore can make categorizations that seem outright metaphorical (e.g., a functional similarity between “climbing the career ladder” and climbing a literal physical ladder).

But these processes can also go awry in disease, Barrett and Miller note. Depression can be seen as a disorder in which the brain imposes overly broad categories, such as “threat” or “criticism” on sensory episodes that don’t have to be perceived that way. By contrast, autism can manifest with features of inadequate compression of incoming sensory signals, not generalizing enough to recognize when a situation is similar enough to a prior one to select the appropriate plan.

Funding to support the paper came from the National Institutes of Health, The U.S. Army Research Institute for the Behavioral and Social Sciences, the Office of Naval Research, the Unlikely Collaborators Foundation, The Freedom Together Foundation, and The Picower Institute for Learning and Memory.

de MIT News https://ift.tt/SreCjuP

Photonics advance could enable compact, high-performance lidar sensors

Lidar systems use pulses of infrared light to measure distance and map a 3D scene with high resolution, allowing autonomous vehicles to rapidly react to obstacles that appear in their path. But traditional lidar sensors are expensive, bulky systems with many moving parts that degrade over time, limiting how the sensors can be deployed.

A new study from MIT researchers could help to enable next-generation lidar sensors that are compact, durable, and have no moving parts. The key advance is a novel design for a silicon-photonics chip, which is a semiconductor device that manipulates light rather than electricity. 

Typically, such silicon-photonics chip-based systems have a restricted field of view, so a silicon-photonics-based lidar would not be able to scan angles in the periphery. Existing workarounds to this problem increase noise and hamper precision.

To avoid these drawbacks, the MIT researchers designed and demonstrated an array of integrated antennas that minimizes unwanted crosstalk between the antennas. Their innovation allows a lidar chip to scan a wider field of view while maintaining low-noise operation compared to other silicon-photonics-based approaches.

This novel demonstration could fuel the development of advanced lidar sensors for demanding applications like autonomous vehicle navigation, aerial surveying, and construction site monitoring.

“The functionality we demonstrated in this work solves a fundamental problem for integrated optical-phased-array technology, enabling future lidar sensors that can achieve significantly higher performance than we could demonstrate previously,” says Jelena Notaros, the Robert J. Shillman Career Development Associate Professor of Electrical Engineering and Computer Science (EECS) at MIT, a member of the Research Laboratory of Electronics, and senior author of a paper on this innovation.

She is joined on the paper by lead author and EECS graduate student Henry Crawford-Eng as well as EECS graduate students Andres Garcia Coleto, Benjamin M. Mazur, Daniel M. DeSantis, and Tal Sneh. The research appears today in Nature Communications.

Adjusting an antenna array

Many traditional lidar systems map a scene using a bulky box that spins to send pulses of light in multiple directions. The light bounces off nearby objects and returns to the sensor, providing data that are used to reconstruct the environment. 

Instead, silicon-photonics-based lidar sensors systematically scan an emitted light beam in multiple directions non-mechanically using a system called an integrated optical phased array (OPA).

Key to an OPA is an array of integrated antennas that have tiny perturbations placed periodically along their length. These corrugations allow the antenna to scatter light from an input source up and out of the photonic chip.

By adjusting the phase of light routed to each antenna, the researchers can change the angle at which the light is emitted out of the array. In this way, they can steer the beam with no moving parts.

But if engineers place the antennas too close together, the antennas will couple with each other and the light they emit will get jumbled. To avoid this, scientists typically space the antennas farther apart, but this also has downsides.

If the antennas are spaced too far apart, the array will emit multiple copies of the light beam at different angles. The researchers can only steer the primary beam so far in either direction until it is undiscernible from its neighboring copies.

“This limits our field of view, so the autonomous vehicle now only knows what is in front of it for a certain angular range,” Garcia Coleto explains.

These beam copies, known as grating lobes, can cause false positives by confusing the sensor. They also waste power.

The MIT researchers solved this problem by designing a set of reduced-crosstalk antennas that can be placed close together without causing a significant coupling effect.

In a standard OPA, all the antennas have the same design, meaning the same arrangement of corrugations. These identical antennas couple very strongly when placed close together.

To address this fundamental roadblock, the MIT researchers designed a set of three antennas with different geometries, varying the width of each antenna and the size and arrangement of corrugations. With varied geometries, each antenna has a different propagation coefficient, which determines how light travels down the antenna.

“Because the antennas have very different propagation coefficients, when we put them close together, essentially each antenna doesn’t ‘see’ the antenna next to it. Therefore, it won’t couple with its neighbor,” Garcia Coleto says. 

A photonic balancing act

But even though the antennas have different propagation coefficients, the researchers still need them to emit light in the same way. 

They achieved this by carefully designing the antennas to meet three parameters. 

First, each antenna must emit the same amount of light. Second, each antenna must emit a beam at the same angle for the same wavelength of light. Third, the emission angle must change uniformly across the array as the researchers steer it.

“We have this challenge where we require the antennas to have different geometries to reduce the crosstalk, but we need to simultaneously design the antennas to have the same emission characteristics. While it is possible to engineer this, it is extremely difficult because, typically, when antennas are designed with different geometries, they tend to behave differently,” Crawford-Eng says.

The researchers first developed the fundamental electromagnetic theory behind how radiative modes couple. They used that theory as a guide to design and simulate their antennas.

Building on those analyses, they fabricated the OPA with reduced-crosstalk antennas spaced significantly closer than they would be in a traditional OPA, then experimentally tested the system.

While a typical OPA would have coupling of about 100 percent in this experiment, their OPA reduced coupling to about 1 percent while generating a single, precise beam. Using this design, they demonstrated accurate beam steering across a wide field of view without any grating lobes. 

In the future, the researchers plan to further improve their technique to enable an even wider field of view. In addition, they are exploring a new potential solution to wide field-of-view functionality that they discovered while developing the underlying theory.

“This work addresses a longstanding challenge in integrated optical phased arrays: simultaneously achieving both a wide field of view, which requires dense antenna spacing, and high beam quality, which requires low crosstalk between neighboring antennas. The authors solve this problem with an elegant antenna design. Their innovation is an important step forward for chip-scale, solid-state beam-steering technology,” says Joyce Poon, professor of electrical and computer engineering at the University of Toronto and director of the Max Planck Institute of Microstructure Physics, who was not involved with this work.

This research was supported, in part, by the Semiconductor Research Corporation, the National Science Foundation, an MIT MathWorks Fellowship, the U.S. Department of War, and the MIT Rolf G. Locher Endowed Fellowship.

de MIT News https://ift.tt/7tfjIZg

Study: Firms often use automation to control certain workers’ wages

When we hear about automation and artificial intelligence replacing jobs, it may seem like a tsunami of technology is going to wipe out workers broadly, in the name of greater efficiency. But a study co-authored by an MIT economist shows markedly different dynamics in the U.S. since 1980. 

Rather than implement automation in pursuit of maximal productivity, firms have often used automation to replace employees who specifically receive a “wage premium,” earning higher salaries than other comparable workers. In practice, that means automation has frequently reduced the earnings of non-college-educated workers who had obtained better salaries than most employees with similar qualifications. 

This finding has at least two big implications. For one thing, automation has affected the growth in U.S. income inequality even more than many observers realize. At the same time, automation has yielded a mediocre productivity boost, plausibly due to the focus of firms on controlling wages rather than finding more tech-driven ways to enhance efficiency and long-term growth.

“There has been an inefficient targeting of automation,” says MIT’s Daron Acemoglu, co-author of a published paper detailing the study’s results. “The higher the wage of the worker in a particular industry or occupation or task, the more attractive automation becomes to firms.” In theory, he notes, firms could automate efficiently. But they have not, by emphasizing it as a tool for shedding salaries, which helps their own internal short-term numbers without building an optimal path for growth.

The study estimates that automation is responsible for 52 percent of the growth in income inequality from 1980 to 2016, and that about 10 percentage points derive specifically from firms replacing workers who had been earning a wage premium. This inefficient targeting of certain employees has offset 60-90 percent of the productivity gains from automation during the time period.

“It’s one of the possible reasons productivity improvements have been relatively muted in the U.S., despite the fact that we’ve had an amazing number of new patents, and an amazing number of new technologies,” Acemoglu says. “Then you look at the productivity statistics, and they are fairly pitiful.”

The paper, “Automation and Rent Dissipation: Implications for Wages, Inequality, and Productivity,” appears in the May print issue of the Quarterly Journal of Economics. The authors are Acemoglu, who is an Institute Professor at MIT; and Pascual Restrepo, an associate professor of economics at Yale University.

Inequality implications

Dating back to the 2010s, Acemoglu and Restrepo have combined to conduct many studies about automation and its effects on employment, wages, productivity, and firm growth. In general, their findings have suggested that the effects of automation on the workforce after 1980 are more significant than many other scholars have believed. 

To conduct the current study, the researchers used data from many sources, including U.S. Census Bureau statistics, data from the bureau’s American Community Survey, industry numbers, and more. Acemoglu and Restrepo analyzed 500 detailed demographic groups, sorted by five levels of education, as well as gender, age, and ethnic background. The study links this information to an analysis of changes in 49 U.S. industries, for a granular look at the way automation affected the workforce. 

Ultimately, the analysis allowed the scholars to estimate not just the overall amount of jobs erased due to automation, but how much of that consisted of firms very specifically trying to remove the wage premium accruing to some of their workers. 

Among other findings, the study shows that within groups of workers affected by automation, the biggest effects occur for workers in the 70th-95th percentile of the salary range, indicating that higher-earning employees bear much of the brunt of this process. 

And as the analysis indicates, about one-fifth of the overall growth in income inequality is attributable to this sole factor.

“I think that is a big number,” says Acemoglu, who shared the 2024 Nobel Prize in economic sciences with his longtime collaborators Simon Johnson of MIT and James Robinson of the University of Chicago.

He adds: “Automation, of course, is an engine of economic growth and we’re going to use it, but it does create very large inequalities between capital and labor, and between different labor groups, and hence it may have been a much bigger contributor to the increase in inequality in the United States over the last several decades.” 

The productivity puzzle

The study also illuminates a basic choice for firm managers, but one that gets overlooked. Imagine a type of automation — call-center technology, for instance — that might actually be inefficient for a business. Even so, firm managers have incentive to adopt it, reduce wages, and oversee a less productive business with increased net profits.

Writ large, some version of this seems to have been happening to the U.S. economy since 1980: Greater profitability is not the same as increased productivity.

“Those two things are different,” says Acemoglu. “You can reduce costs while reducing productivity.” 

Indeed, the current study by Acemoglu and Restrepo calls to mind an observation by the late MIT economist Robert M. Solow, who in 1987 wrote, “You can see the computer age everywhere but in the productivity statistics.” 

In that vein, Acemoglu observes, “If managers can reduce productivity by 1 percent but increase profits, many of them might be happy with that. It depends on their priorities and values. So the other important implication of our paper is that good automation at the margins is being bundled with not-so-good automation.” 

To be clear, the study does not necessarily imply that less automation is always better. Certain types of automation can boost productivity and feed a virtuous cycle in which a firm makes more money and hires more workers. 

But currently, Acemoglu believes, the complexities of automation are not yet recognized clearly enough. Perhaps seeing the broad historical pattern of U.S. automation, since 1980, will help people better grasp the tradeoffs involved — and not just economists, but firm managers, workers, and technologists. 

“The important thing is whether it becomes incorporated into people’s thinking and where we land in terms of the overall holistic assessment of automation, in terms of inequality, productivity and labor market effects,” Acemoglu says. “So we hope this study moves the dial there.”

Or, as he concludes, “We could be missing out on potentially even better productivity gains by calibrating the type and extent of automation more carefully, and in a more productivity-enhancing way. It’s all a choice, 100 percent.”

de MIT News https://ift.tt/zJRZ75N

MIT BrainTrust supports neighbors living with brain injuries

Since 1998, members of MIT’s BrainTrust club have helped Boston-area residents with brain injuries or other neurological disorders through their buddy program. The organization’s members also visit patients in nursing homes suffering from neurological issues.

BrainTrust is one of the founding chapters of Synapse National, an organization created by MIT alumna Alissa Totman ’13. Synapse’s goal is to provide social support for individuals living with brain injuries and to educate and inspire student leaders in the field of brain injury.

“Learning directly from individuals who had experienced brain injury during my time in BrainTrust gave me an appreciation of the gaps in resources and opportunities for improvement in brain injury care, which ultimately motivated me to pursue a career in brain injury medicine. My experience in BrainTrust continues to shape my approach to patient care and my professional goal of improving access to specialized care for individuals with brain injury by serving as a consulting provider in the acute care hospital, as well as by training the next generation of leaders in the field,” says Totman.

The club’s president, junior Karie Shen, who is pursuing a double major in biology (Course 7) and brain and cognitive science (Course 9), says, “BrainTrust is a student-run service organization that provides support for individuals with brain injury and other neurological disorders. I joined BrainTrust because it seemed like the perfect intersection of community service and neuroscience, and I care about these two things deeply.”

BrainTrust volunteers participate in training and then are paired with a local buddy who has experienced a brain injury. Members can also spend time on the weekends with patients in nursing homes who have dementia, Alzheimer’s disease, or who have had a stroke.

Shen, along with Elizabeth Zhang, president of the MIT Pre-Med Society, recently developed a program that allows BrainTrust members to visit patients in hospice. “It’s an experience that is deeply valuable for students. We work through a third-party organization called Compassus. Because the pairing process is HIPAA-protected, our role as BrainTrust executive members is to recruit students and connect them with the hospice volunteer coordinator for training. We also provide funding for transportation, generously supported by the UA Community Service Committee,” says Shen.

Shen, who plans to go to medical school and specialize in neurology, neuro-oncology, or geriatric medicine when she completes her degree, finds the experience rewarding, at times difficult, but also offers a glimpse into the reality of working with people with brain injuries.

“Visiting the people in hospice or a nursing home is hard. I’ve seen residents cry for no apparent reason that the nurses or I can understand. But I have also come to understand that caring for a patient’s quality of life and dignity is equally important. What I came to realize is that my presence itself mattered. That perspective has shaped how I think about the kind of physician I want to become,” says Shen.

First-year student Jordan Lacsamana heard about the club during Campus Preview Weekend and was immediately interested. Lacsamana, who will major in brain and cognitive sciences, is a volunteer in the Buddy Program and meets with her buddy at least once a month.

“I joined the club because it aligned with my interests academically, but I also wanted to support someone in the Boston community. I’m pre-med, and I’m interested in surgery, possibly neurosurgery or cardiovascular surgery. But I also think it’s nice to have someone outside of MIT to talk with. It’s great to learn more about them and have that one-on-one friendship, which really is the goal,” says Lacsamana.

Lacsamana says she enjoys spending time with Amanda, her buddy, and exploring Boston and Harvard Square, meeting for coffee or meals, and getting as much out of the relationship as Amanda does.

“I see her as a mentor because coming to Boston from Dallas was such a big change, so I’ve also been able to look to her for advice. But I think one of the great things about the program is that you get to learn more about them as an individual, instead of seeing them as just a person with an injury,” says Lacsamana.

“Many of our brain injury buddies simply enjoy being around students, staying connected to what we are learning and doing. Some have been with the club for years, even upwards of a decade, and still keep up with former student members long after they graduate. It is really wonderful to see how BrainTrust has created this web of friendships between people who would otherwise never have met,” says Shen.

“Amanda has stayed in touch with her former buddy since she graduated from MIT and is going to her wedding,” says Lacsamana. “I think it’s a testament to how amazing this program is at forming those connections.”

MIT students who seek real-world opportunities in fields such as cognitive science, health care, medicine, and cognitive/neurological prosthetics, or who want to help a local resident, can join BrainTrust. Email braintrust-exec@mit.edu for more information.

de MIT News https://ift.tt/91LoZHU

Method for stress-testing cloud computing algorithms helps avoid network failures

Researchers from MIT and elsewhere have developed a more user-friendly and efficient method to help networking engineers identify potential system failures before they cause major problems, like a cloud service outage that leaves millions of users unable to access applications. 

The technique uncovers hidden blind spots that might cause a shortcut algorithm to fail unexpectedly when it is deployed. 

This new approach can identify worse-case scenarios that an engineer might miss if they use a traditional method that compares an algorithm against a set of human-designed past test cases. It is also less labor-intensive than other verification tools that require engineers to rewrite an algorithm in a complex mathematical code each time they want to test it.

Instead of needing a mathematical reformulation, the new method reads the algorithm’s source code directly and automatically searches for worse-case scenarios that lead to the highest level of underperformance.

By helping engineers quickly and easily stress-test a networking algorithm before deployment, the method could catch failure modes that might otherwise only appear in a real outage. The technique could also be used to analyze the risks of deploying AI-generated code.

“We need to have good tools to measure the worse-case scenario performance of our algorithms so we know what could happen before we put them into production. This is an easy-to-use tool that can be plugged into current systems so we can find the best algorithm to use and ensure the worse-case scenarios are identified in advance,” says Pantea Karimi, an electrical engineering and computer science (EECS) graduate student and lead author of a paper on this new technique. 

She is joined on the paper by senior authors Mohammad Alizadeh, an associate professor of EECS and a member of the Computer Science and Artificial Intelligence Laboratory (CSAIL); and Behnaz Arzani, a principal researcher at Microsoft Research; along with Ryan Beckett, Siva Kesava Reddy Karkarla, and Pooria Namyar, researchers at Microsoft Research; and Santiago Segarra, a professor at Rice University. The research will be presented at the USENIX Symposium on Networked Systems Design and Implementation. 

Assessing algorithms

In large systems like cloud servers, the tried-and-true algorithms that route data from one place to another or are often too computationally intensive to run in a feasible amount of time. 

So, engineers and researchers develop suboptimal algorithms called heuristics that can run much faster. However, there could be unexpected but plausible circumstances that will cause a heuristic to underperform or fail when deployed.

A heuristic can route millions of data requests across a cloud network in seconds, but under the wrong conditions — like an unusual traffic pattern or a sudden spike in demand — the shortcut can break down in ways the designer never anticipated.

When these problems occur, a company may have no choice but to drop some requests that can’t be processed. 

The firm could also deliberately allocate more resources in advance to head-off a potential disaster, leading to higher overall costs and wasted electricity from underutilization.

“This is really bad for a company because, either way, they are going to lose a lot of money. If this particular scenario hasn’t happened before and was never tested, how would a developer know in advance before it happens?” Karimi says.

Stress-testing heuristics typically involves running a new algorithm in simulation using a set of human-designed test cases and manually comparing the performance with a previous algorithm. But this is time-consuming and can leave blind spots if an engineer doesn’t know to test for certain situations.

Alternatively, engineers could use a verification tool to evaluate the performance of their heuristic more systematically. However, these tools require the engineer to encode the algorithm into a complex, mathematical formula that can take days to flesh out. The process, which doesn’t work for every type of heuristic, must be repeated each time the engineer changes the code.

Instead, the researchers developed a more user-friendly and efficient verification tool, called MetaEase, that analyzes the heuristic’s existing implementation code directly to identify the biggest risks of deploying it.

“This would reduce the friction of using these heuristic analysis tools,” Karimi says.

She began this work during an internship at Microsoft Research, where the team previously developed MetaOpt, a heuristic analyzer that requires engineers to rewrite their algorithms as formal optimization models. MetaEase grew out of the desire to remove that barrier.

Maximizing the gap

MetaEase is driven by two key innovations. First, it uses a technique called symbolic execution to map out the different decision points in the heuristic's code. These are places where the algorithm might behave differently depending on the input.

This technique produces a set of representative starting points, each corresponding to a distinct behavior the heuristic could exhibit.

Second, from these starting points, MetaEase utilizes a guided search to systematically move toward inputs that make the heuristic perform as poorly as possible, compared to the optimal algorithm.

In machine learning, for instance, an input could be a set of user queries to an AI chatbot at a given time.

“In this way, we have exploited every possible heuristic behavior and used special techniques to move in the direction where we think the performance gap is going to increase,” Karimi explains.

In the end, MetaEase identifies the input that maximizes the performance gap between the heuristic and an optimal benchmark.

With this information, a heuristic developer could inspect the input to understand what went wrong and incorporate safeguards that will prevent the problem from happening during deployment.

In simulated experiments, MetaEase often identified inputs with larger performance gaps than traditional methods — pinpointing more catastrophic worse-case scenarios. And it did so much more efficiently. 

It was also able to analyze a recent networking heuristic that no state-of-the-art method could handle.

In the future, the researchers want to enhance MetaEase so it can process additional types of types of data, like categorical inputs. They also want to improve the scalability of their method and adapt MetaEase to evaluate more complex heuristics.

“Reasoning about the worst-case performance of deployed heuristics is a hard and longstanding problem. MetaEase makes tangible progress by analyzing heuristics directly from source code, eliminating the need for formal models that have historically limited who can use such analysis tools. I was pleasantly surprised that it handles non-convex and randomized heuristics by combining symbolic execution with gradient-based search in a practical and effective way,” says Ratul Mahajan of the University of Washington Paul G. Allen School of Computer Science and Engineering, who was not involved with this research.

This research was funded, in part, by a Microsoft Research internship and the U.S. National Science Foundation (NSF).

de MIT News https://ift.tt/whMjio8

MIT marks first Robert R. Taylor Day with Tuskegee University

On April 10, MIT marked its first official Robert R. Taylor Day with a program centered on the life and work of Robert Robinson Taylor (Class of 1892), the Institute’s first Black graduate and the first academically trained Black architect in the United States.

After graduating from MIT, Taylor joined Tuskegee Institute (now Tuskegee University), where he designed campus buildings, developed a curriculum, and helped establish an approach to architectural education grounded in making and community life — an orientation that continues to shape the relationship between MIT and Tuskegee today. 

Taylor returned to MIT on April 10, 1911, to speak at the 50th anniversary of the Institute’s founding — the date now observed as Robert R. Taylor Day. Reflecting on his education, he credited MIT with the “methods and plans” he carried to Tuskegee Institute. “Certainly the spirit,” he said, was found “in the love of doing things correctly, of putting logical ways of thinking into the humblest task … to build up the immediate community in which the persons live.”

One hundred fifteen years later, at the MIT Museum, students and faculty gathered around Taylor’s original thesis, “A Soldiers Home.” The work was presented alongside archival materials from Taylor’s time at MIT by Jonathan Duval, assistant curator of architecture and design. Rather than framing Taylor as a distant historical figure, the encounter with the work itself — its drawings, assumptions, and ambitions — set the terms for the day, bringing forward not only his accomplishments but the ideas and methods that continue to inform teaching and collaboration today. Attendees then gathered for a lunch-and-learn session including a hybrid panel involving MIT and Tuskegee University faculty. 

“It is so important to continue to develop the MIT-Tuskegee relationship begun by Robert R. Taylor,” says Kwesi Daniels, associate professor and head of the architecture department at Tuskegee University. “MIT students are provided an opportunity to experience the campus Taylor designed and his ethos of social architecture. For the Tuskegee students, they are able to appreciate the foundation Taylor received at MIT. The engagement epitomizes the ‘mind and hand’ philosophy of MIT and the head, hand, heart philosophy of Tuskegee.”

An ongoing exchange

Student and faculty exchanges, launched by the architecture departments at both institutions, have extended these connections in recent years. MIT students travel to Tuskegee for work in historic preservation and community engagement, sampling Daniels’ scanning and drone equipment, while Tuskegee students come to MIT to engage with digital fabrication and entrepreneurship.

For Nicholas de Monchaux, professor and head of the Department of Architecture at MIT, the relationship reflects continuity. “We are not uniting. We’re reuniting,” he says. “This year’s celebration should really be seen as the kickoff of a year of reflecting on Robert Taylor’s legacy and imagining what the day, and his legacy, can become over time.”

The day’s program — the vision for which originally emerged from a suggestion made by MIT literature professor Joshua Bennett during a meeting at Tuskegee with de Monchaux, Daniels, and Tuskegee President Mark Brown — moved into a broader effort among faculty and collaborators across architecture, history, and the humanities. As Bennett put it, “The primary aim of Robert R. Taylor Day is to lift up not only Taylor’s accomplishments, but his ideas — and the fact that his ideas live on in those of us who have inherited his legacy.”

That emphasis is also visible in the dedicated coursework and research that has accompanied the exchange since 2022. In class 4.s12 (Brick x Brick: Drawing a Particular Survey), taught by Carrie Norman, assistant professor in architecture at MIT, students document buildings on the Tuskegee campus through measured drawings and archival interpretation. Working from limited historical material, they reconstruct both form and intent.

“My role has been to structure this work pedagogically,” Norman says, “guiding students in methods of close looking, measured drawing, and archival interpretation.” She describes Taylor’s work as “an ongoing research agenda,” adding that “the broader aim is not only to deepen engagement with Taylor’s legacy, but to build on it through new forms of design research.”

Related work has contributed to a recent exhibition on the Tuskegee Chapel at the National Building Museum, curated by Helen Bechtel of the Yale School of Architecture. Building on research conducted in Norman’s course, students developed large-scale models that form part of the exhibition. New 3D fabrications use a limited set of archival materials to reconstruct the chapel originally designed by Taylor as the first electrified building in Alabama’s Macon County, which was destroyed by fire in 1957.

Looking ahead

Timothy Hyde, professor in the MIT Department of Architecture, has also been involved in the ongoing MIT–Tuskegee collaboration and in efforts to situate Taylor’s work within a broader historical context. He notes that Taylor’s training at MIT helped shape the curriculum he later developed at Tuskegee. “The other influence I would like to mention is the city of Boston itself,” Hyde adds. “Boston was a prosperous city with a wealth of civic architecture that Taylor would have seen and studied.” 

A documentary project on Taylor’s life, supported by the MIT Human Insight Collaborative and led by Hyde and historian Christopher Capozzola, senior associate dean for MIT Open Learning, is currently in development.

For some students, these encounters shape longer trajectories. As an undergraduate at Tuskegee, Myles Sampson participated in the MIT Summer Research Program (MSRP), where he began to connect architecture with a growing interest in computation. He later enrolled in MIT’s Master of Science in Architecture Studies (SMArchS) computation program, working with Professor Larry Sass, who introduced him to robotic fabrication.

“I never looked back,” Sampson says. “Without that hands-on research experience, I would never have looked past contemporary architectural practice.” He is now pursuing a doctorate in computational design at Carnegie Mellon University, focused on the role of automation in architecture and construction.

Sampson contributed significant work to the National Building Museum’s exhibition. His installation, Brick Parable, brings together historical reference and robotic construction. As de Monchaux notes, the project reflects the long arc of Taylor’s legacy: “bricks were fired by students as part of Taylor’s training program … Myles [Sampson]’s piece, made with a robotic assembly of bricks, explores the architectural idea of the chapel in contemporary form.”

For Daniels, the continued circulation of students between the two institutions remains central. Viewing Taylor’s thesis in particular offers a shared point of reference. “Whether the student is from Tuskegee or MIT, they are able to appreciate the quality of work Taylor completed as a student,” he says, “and how he built on that work by creating a college campus, beginning at age 25.”

Across these activities, Taylor’s work is approached not as a fixed legacy, but as a set of methods and commitments that continue to be tested. As Catherine Armwood, dean of Tuskegee University Robert R. Taylor School of Architecture and Construction Science, describes it: “While our students leverage [the design and entrepreneurship program] MITdesignX to turn architectural concepts into social enterprises through advanced fabrication and venture mentorship, MIT students come to Tuskegee for an immersion in historic preservation. By surveying buildings handcrafted by our founding students, they learn a legacy of self-reliance and community impact that can’t be found anywhere else,” Armwood says. “Together, we are bridging technical innovation with deep-rooted heritage to train a new generation of visionary leaders.” 

de MIT News https://ift.tt/b8pq1xv