Artificial Intelligence

Decades of technological change have polarized the earnings of the American workforce, helping highly educated white-collar workers thrive, while hollowing out the middle class. Yet present-day advances like robots and artificial intelligence do not spell doom for middle-tier or lower-wage workers, since innovations create jobs as well. With better policies in place, more people could enjoy good careers even as new technology transforms workplaces.
That’s the conclusion of the final report from MIT’s Task Force on the Work of the Future, which summarizes over two years of research on technology and jobs. The report, “The Work of the Future: Building Better Jobs in an Age of Intelligent Machines,” was released today, and the task force is hosting an online conference on Wednesday, the “AI & the Future of Work Congress.”
At the core of the task force’s findings: A robot-driven jobs apocalypse is not on the immediate horizon. As technology takes jobs away, it provides new opportunities; about 63 percent of jobs performed in 2018 did not exist in 1940. Rather than a robot revolution in the workplace, we are witnessing a gradual tech evolution. At issue is how to improve the quality of jobs, particularly for middle- and lower-wage workers, and ensure there is greater shared prosperity than the U.S. has seen in recent decades.
“The sky is not falling, but it is slowly lowering,” says David Autor, the Ford Professor of Economics at MIT, associate head of MIT’s Department of Economics, and a co-chair of the task force. “We need to respond. The world is gradually changing in very important ways, and if we just keep going in the direction we’re going, it is going to produce bad outcomes.”
That starts with a realistic understanding of technological change, say the task force leaders.
The task force aimed “to move past the hype about what [technologies] might be here, and now we’re looking at what can we feasibly do to move things forward for workers,” says Elisabeth Beck Reynolds, executive director of the task force as well as executive director of the MIT Industrial Performance Center. “We looked across a range of industries and examined the numerous factors — social, cognitive, organizational, economic — that shape how firms adopt technology.”
“We want to inject into the public discourse a more nuanced way of talking about technology and work,” adds David Mindell, task force co-chair, professor of aeronautics and astronautics, and the Dibner Professor of the History of Engineering and Manufacturing at MIT. “It’s not that the robots are coming tomorrow and there’s nothing we can do about it. Technology is an aggregate of human choices.”
The report also addresses why Americans may be concerned about work and the future. It states: “Where innovation fails to drive opportunity, it generates a palpable fear of the future: the suspicion that technological progress will make the country wealthier while threatening the people’s livelihoods. This fear exacts a high price: political and regional divisions, distrust of institutions, and mistrust of innovation itself. The last four decades of economic history give credence to that fear.”
“Automation is transforming our work, our lives, our society,” says MIT President L. Rafael Reif, who initiated the formation of the task force in 2017. “Fortunately, the harsh societal consequences that concern us all are not inevitable. How we design tomorrow’s technologies, and the policies and practices we build around them, will profoundly shape their impact.”
Reif adds: “Getting this right is among the most important and inspiring challenges of our time — and it should be a priority for everyone who hopes to enjoy the benefits of a society that’s healthy and stable, because it offers opportunity for all.”
Six big conclusions
The task force, an Institute-wide group of scholars and researchers, spent over two years studying work and technology in depth. The final report presents six overarching conclusions and a set of policy recommendations. The conclusions:
1) Technological change is simultaneously replacing existing work and creating new work. It is not eliminating work altogether.
Over the last several decades, technology has significantly changed many workplaces, especially through digitization and automation, which have replaced clerical, administrative, and assembly-line workers across the country. But the overall percentage of adults in paid employment has largely risen for over a century. In theory, the report states, there is “no instrinsic conflict between technological change, full employment, and rising earnings.”
In practice, however, technology has polarized the economy. White-collar workers — in medicine, marketing, design, research, and more — have become more productive and richer, while middle-tier workers have lost out. Meanwhile, there has been growth in lower-paying service-industry jobs where digitization has little impact — such as food services, janitors, and drivers. Since 1978, aggregate U.S. productivity has risen by 66 percent, while compensation for production and nonsupervisory workers has risen by only 10 percent. Wage gaps also exist by race and gender, and cities do not provide the “escalator” to the middle class they once did.
While innovations have replaced many receptionists, clerks, and assembly-line workers, they have simultaneously created new occupations. Since the middle of the 20th century, the U.S. has seen major growth in the computer industry, renewable energy, medical specialties, and many areas of design, engineering, marketing, and health care. These industries can support many middle-income jobs as well — while the services sector keeps growing.
As the task force leaders state in the report, “The dynamic interplay among task automation, innovation, and new work creation, while always disruptive, is a primary wellspring of rising productivity. Innovation improves the quantity, quality, and variety of work that a worker can accomplish in a given time. This rising productivity, in turn, enables improving living standards and the flourishing of human endeavors.”
However, a bit ruefully, the authors also note that “in what should be a virtuous cycle, rising productivity provides society with the resources to invest in those whose livelihoods are disrupted by the changing structure of work.”
But this has not come to pass, as the distribution of value from these jobs has been lopsided. In the U.S., lower-skill jobs only pay 79 percent as much when compared to Canada, 74 percent compared to the U.K., and 57 percent compared to Germany.
“People understand that automation can make the country richer and make them poorer, and that they’re not sharing in those gains,” Autor says. “We think that can be fixed.”
2) Momentous impacts of technological change are unfolding gradually.
Time and again, media coverage about technology and jobs focuses on dramatic scenarios in which robots usurp people, and we face a future without work.
But this picture elides a basic point: Technologies mimicking human actions are difficult to build, and expensive. It is generally cheaper to simply hire people for those tasks. On the other hand, technologies that augment human abilities — like tools that let doctors make diagnoses — help those workers become more productive. Apart from clerical and assembly-line jobs, many technologies exist in concert with workers, not as a substitute for them.
Thus workplace technology usually involves “augmentation tasks more than replacement tasks,” Mindell says. The task force report surveys technology adoption in industries including insurance, health care, manufacturing, and autonomous vehicles, finding growth in “narrow” AI systems that complement workers. Meanwhile, technologists are working on difficult problems like better robotic dexterity, which could lead to more direct replacement of workers, but such advances at a high level are further off in the future.
“That’s what technological adoption looks like,” Mindell says. “It’s uneven, it’s lumpy, it goes in fits and starts.” The key question is how innovators at MIT and elsewhere can shape new technology to broad social benefit.
3) Rising labor productivity has not translated into broad increases in incomes because societal institutions and labor market policies have fallen into disrepair.
While the U.S. has witnessed a lot of technological innovation in recent decades, it has not seen as much policy innovation, particularly on behalf of workers. The polarizing effects of technology on jobs would be lessened if middle- and lower-income workers had relatively better support in other ways. Instead, in terms of pay, working environment, termination notice time, paid vacation time, sick time, and family leave, “less-educated and low-paid U.S. workers fare worse than comparable workers in other wealthy industrialized nations,” the report notes. The adjusted gross hourly earnings of lower-skill workers in the U.S. in 2015 averaged $10.33, compared to $24.28 in Denmark, $18.18 in Germany, and $17.61 in Australia.
“It’s untenable that the labor market has this growing gulf without shared prosperity,” Autor says. “We need to restore the synergy between rising productivity and improvements in labor market opportunity.” He adds: “We’ve had real institutional failure, and it’s within our hands to change it. … That includes worker voice, minimum wages, portable benefits, and incentives that cause companies to invest in workers.”
Looking ahead, the report cautions, “If those technologies deploy into the labor institutions of today, which were designed for the last century, we will see similar effects to recent decades: downward pressure on wages, skills, and benefits, and an increasingly bifurcated labor market.” The task force argues instead for institutional innovations that complement technological change.
4) Improving the quality of jobs requires innovation in labor market institutions. 
The task force contends the U.S. needs to modernize labor policies on several fronts, including restoring the federal minimum wage to a reasonable percentage of the national median wage and, crucially, indexing it to inflation. 
The report also suggests upgrading unemployment insurance in several ways, including: using very recent earnings to determine eligibility or linking eligibility to hours worked, not earnings; making it easier to receive partial benefits in case of events like loss of a second job; and dropping the requirement that people need to seek full-time work to receive benefits, since so many people hold part-time positions. 
The report also observes that U.S. collective bargaining law and processes are antiquated. The authors argue that workers need better protection of their current collective bargaining rights; new forms of workplace representation beyond traditional unions; and legal protections allowing groups to organize that include home-care workers, farmworkers, and independent contractors.
5) Fostering opportunity and economic mobility necessitates cultivating and refreshing worker skills.
Technological advancement may often be incremental, but changes happen often enough that workers’ skills and career paths can become obsolete. The report emphasizes that U.S. workers need more opportunities to add new skills — whether through the community college system, online education, company-based retraining, or other means.  
The report calls for making ongoing skills development accessible, engaging, and cost-effective. This requires buttressing what already works, while advancing new tools: blended online and in-person offerings, machine-supervised learning, and augmented and virtual reality learning environments.
The greatest needs are among workers without four-year college degrees. “We need to focus on those who are between high school and the four-year degree,” Reynolds says. “There should be pathways for those people to increase their skill set and make it meaningful to the labor market. We really need a shift that makes this a high priority.”
6) Investing in innovation will drive new job creation, speed growth, and meet rising competitive challenges.
The rate of new-job creation over the last century is heavily driven by technological innovation, the report notes, with a considerable portion of that stemming from federal investment in R&D, which has helped produce many forms of computing and medical advances, among other things. As of 2015, the U.S. invested 2.7 percent of its GDP in R&D, compared to 2.9 percent in Germany and 2.1 percent in China. But the public share of that R&D investment has fallen from 40 percent in 1985 to 25 percent in 2015. The task force calls for a recommitment to this federal support.
“Innovation has a key role in job creation and growth,” Autor says.
Given the significance of innovation to job and wealth creation, the report calls for increased overall federal research funding; targeted assistance that helps small- and medium-sized businesses adopt technology; policies creating a wider geographical spread of innovation in the U.S.; and policies that enhance investment in workers, not just capital, including the elimination of accelerated capital depreciation claims, and an employer training tax credit that functions like the R&D tax credit.
Global issues, U.S. suggestions
In addition to Reynolds, Autor, and Mindell, MIT’s Task Force on the Work of the Future consisted of a group of 18 MIT professors representing all five Institute schools and the MIT Schwarzman College of Computing; a 22-person advisory board drawn from the ranks of industry leaders, former government officials, and academia; a 14-person research board of scholars; and over 20 graduate students. The task force also consulted with business executives, labor leaders, and community college leaders, among others. The final document includes case studies from specific firms and sectors as well, and the Task Force is publishing nearly two dozen research briefs that go into the primary research in more detail. 
The task force observed global patterns at play in the way technology is adopted and diffused through the workplace, although its recommendations are focused on U.S. policy issues.
“While our report is very geared toward the U.S. in policy terms, it clearly is speaking to a lot of trends and issues that exist globally,” Reynolds said. “The message is not just for the U.S. Many of the challenges we outline are found in other countries too, albeit to lesser degrees. As we wrote in the report, ‘the central challenge ahead, indeed the work of the future, is to advance labor market opportunity to meet, complement, and shape technological innovations.’”
The task force intends to circulate ideas from the report among policymakers and politicians, corporate leaders and other business managers, and researchers, as well as anyone with an interest in the condition of work in the 21st century.
“I hope people are receptive,” Reynolds adds. “We have made forceful recommendations that tie together different policy areas — skills, job quality, and innovation. These issues are critical, particularly as we think about recovery and rebuilding in the age of Covid-19. I hope our message will be picked up by both the public sector and private sector leaders, because both of those are essential to forge the path forward.” More

As a student in the MIT Laboratory for Information and Decision Systems (LIDS) and in the MIT Institute for Data, Systems, and Society (IDSS), Manon Revel has been investigating how advertising in online publications affects trust in journalism. The topic area is familiar for Revel; growing up, she wanted to be a journalist to follow in her father’s footsteps. As a teenager, Revel was assigned her first reporting gig covering a holiday fireworks display for the local radio station near her home on the Atlantic coast of France. Revel researched every one of the dozens of types of fireworks that would be launched and spent hours interviewing the person who would launch them before reporting live on the night. Revel continued reporting for radio and print (she started a high school newspaper that is still active), using the same thorough approach to prepare for each interview.
“I wanted to ask interesting people questions that they had never been asked before, so I read everything I could on them beforehand and then I tried to learn something new,” Revel says.
However, in France, students must narrow in on a field of study by the end of their secondary schooling, and Revel ultimately chose to focus on science — a field in which her penchant for meticulous research and love of discovery would be well served. She completed a bachelor’s and master’s degree in engineering and applied mathematics at École Centrale Paris while pursuing other interests, including journalism, on the side. As she considered what to do next, Revel sought out a more interdisciplinary opportunity, and discovered that at MIT — in particular at LIDS, and in the Master’s Program in Technology and Policy (TPP) at IDSS.
“I always felt like I had these different lives. Coming to MIT was the first time I felt that my interests could finally be put together, and that I could work on everything in unison,” Revel says.
Revel only intended to come for a master’s degree, but she loved the environment and the people so much that she stayed on for a PhD. She joined the new Social and Engineering Systems (SES) doctoral program within IDSS, which combines tools and methods from statistics and information sciences, engineering, and social sciences to tackle significant societal problems. It was the exact sort of interdisciplinary opportunity Revel had been seeking. She works with four academic advisors: professors Ali Jadbabaie in the Department of Civil and Environmental Engineering, Dean Eckles in the MIT Sloan School of Management, Adam Berinsky in the Department Political Science, and IDSS Director Munther Dahleh in the Department of Electrical Engineering and Computer Science. Though operating in the middle of their fields has sometimes been challenging — like having to speak four different languages, Revel says — she’s been thrilled with the opportunities to combine the approaches of disparate disciplines in order to build on all of them and push beyond the boundaries of previous research.
The timing of Revel’s arrival, shortly after the 2016 United States presidential election, was perfect for a deep dive into journalism from an analytical perspective. Current events had underscored the importance of understanding how information is, for better and worse, disseminated in the digital era and made sense of by the people receiving it. Revel’s main project, together with a fellow graduate student in LIDS, Amir Tohidi, has been investigating the effect of “clickbait” ads on reader trust.
As publications have struggled with online revenue models, many have come to increasingly rely on “native ads” that are designed to blend in with news stories and are often supplied by a third-party network. The majority of these ads are clickbait, which Revel and her colleagues define as “something (such as a headline) designed to make readers want to click on a hyperlink, especially when the link leads to content of dubious value or interest.” Revel wanted to quantify exactly how prevalent clickbait is, so she set up a Bayesian text classification program to recognize it. The AI was trained off of two testing sets labelled by users on Amazon’s Mechanical Turk service. Then, Revel scraped over 1.4 million ads from the time period 2016-19, and, using her text classification program, found that more than 80 percent were clickbait. Next, Revel ran two large-scale randomized experiments, which showed that one-time exposure to clickbait near an article could negatively impact readers’ trust in the article and publication. That effect was driven by medium-familiarity publications — defined as outlets recognized by 25-50 percent of the audience. On the contrary, the ads’ effect on the most well-known outlets, like CNN and Fox News, was found to be null.
Revel hopes that the project will raise awareness among journalism publications of the risk of losing readership in the long term in order to reap the short-term financial rewards of native advertising. She acknowledges that there’s not an easy solution, as publishers may not be able to afford losing the revenue. However, reader trust in high-quality journalism is a crucial issue in a time of rampant “information disorder,” which is the umbrella term that researchers like Revel use to describe the combination of intentionally or maliciously false information, unwittingly shared misinformation, and true but harmful information that spreads like wildfire through the internet.
Recently, Revel has been working on several other projects that combine approaches across disciplines to quantify and decipher human behavior, with a focus on understanding the ways that people make sense of information and how that impacts their voting decisions. She has been working with sociologists on a project about how audiences’ perception of the news resonates with media coverage. Another project is modeling a voting system where votes may be delegated to better-informed acquaintances.
One could imagine that so much time spent thinking about Revel’s chosen research topics — politics, journalism, untruths — could be draining, but Revel tries not to get drawn into the strong emotions that these subjects can provoke. She sees the goal of her work as understanding, rather than judging, how people think and behave. At the same time, in an occupation focused on information disorder, perhaps it is no surprise that one of the things Revel appreciates most about her colleagues is their authenticity. This, along with their willingness to admit what they don’t know, made it easy to stay on at LIDS and IDSS after her master’s, she says. It’s a wonderful environment in which Revel can do what she has always loved, which is to learn everything she can about a subject — whether fireworks, clickbait, or human behavior — and, now aided by the many scientific approaches she has acquired, try to discover something new. More

Deep learning is everywhere. This branch of artificial intelligence curates your social media and serves your Google search results. Soon, deep learning could also check your vitals or set your thermostat. MIT researchers have developed a system that could bring deep learning neural networks to new — and much smaller — places, like the tiny computer chips in wearable medical devices, household appliances, and the 250 billion other objects that constitute the “internet of things” (IoT).
The system, called MCUNet, designs compact neural networks that deliver unprecedented speed and accuracy for deep learning on IoT devices, despite limited memory and processing power. The technology could facilitate the expansion of the IoT universe while saving energy and improving data security.
The research will be presented at next month’s Conference on Neural Information Processing Systems. The lead author is Ji Lin, a PhD student in Song Han’s lab in MIT’s Department of Electrical Engineering and Computer Science. Co-authors include Han and Yujun Lin of MIT, Wei-Ming Chen of MIT and National University Taiwan, and John Cohn and Chuang Gan of the MIT-IBM Watson AI Lab.
The Internet of Things
The IoT was born in the early 1980s. Grad students at Carnegie Mellon University, including Mike Kazar ’78, connected a Cola-Cola machine to the internet. The group’s motivation was simple: laziness. They wanted to use their computers to confirm the machine was stocked before trekking from their office to make a purchase. It was the world’s first internet-connected appliance. “This was pretty much treated as the punchline of a joke,” says Kazar, now a Microsoft engineer. “No one expected billions of devices on the internet.”
Since that Coke machine, everyday objects have become increasingly networked into the growing IoT. That includes everything from wearable heart monitors to smart fridges that tell you when you’re low on milk. IoT devices often run on microcontrollers — simple computer chips with no operating system, minimal processing power, and less than one thousandth of the memory of a typical smartphone. So pattern-recognition tasks like deep learning are difficult to run locally on IoT devices. For complex analysis, IoT-collected data is often sent to the cloud, making it vulnerable to hacking.
“How do we deploy neural nets directly on these tiny devices? It’s a new research area that’s getting very hot,” says Han. “Companies like Google and ARM are all working in this direction.” Han is too.
With MCUNet, Han’s group codesigned two components needed for “tiny deep learning” — the operation of neural networks on microcontrollers. One component is TinyEngine, an inference engine that directs resource management, akin to an operating system. TinyEngine is optimized to run a particular neural network structure, which is selected by MCUNet’s other component: TinyNAS, a neural architecture search algorithm.
System-algorithm codesign
Designing a deep network for microcontrollers isn’t easy. Existing neural architecture search techniques start with a big pool of possible network structures based on a predefined template, then they gradually find the one with high accuracy and low cost. While the method works, it’s not the most efficient. “It can work pretty well for GPUs or smartphones,” says Lin. “But it’s been difficult to directly apply these techniques to tiny microcontrollers, because they are too small.”
So Lin developed TinyNAS, a neural architecture search method that creates custom-sized networks. “We have a lot of microcontrollers that come with different power capacities and different memory sizes,” says Lin. “So we developed the algorithm [TinyNAS] to optimize the search space for different microcontrollers.” The customized nature of TinyNAS means it can generate compact neural networks with the best possible performance for a given microcontroller — with no unnecessary parameters. “Then we deliver the final, efficient model to the microcontroller,” say Lin.
To run that tiny neural network, a microcontroller also needs a lean inference engine. A typical inference engine carries some dead weight — instructions for tasks it may rarely run. The extra code poses no problem for a laptop or smartphone, but it could easily overwhelm a microcontroller. “It doesn’t have off-chip memory, and it doesn’t have a disk,” says Han. “Everything put together is just one megabyte of flash, so we have to really carefully manage such a small resource.” Cue TinyEngine.
The researchers developed their inference engine in conjunction with TinyNAS. TinyEngine generates the essential code necessary to run TinyNAS’ customized neural network. Any deadweight code is discarded, which cuts down on compile-time. “We keep only what we need,” says Han. “And since we designed the neural network, we know exactly what we need. That’s the advantage of system-algorithm codesign.” In the group’s tests of TinyEngine, the size of the compiled binary code was between 1.9 and five times smaller than comparable microcontroller inference engines from Google and ARM. TinyEngine also contains innovations that reduce runtime, including in-place depth-wise convolution, which cuts peak memory usage nearly in half. After codesigning TinyNAS and TinyEngine, Han’s team put MCUNet to the test.
MCUNet’s first challenge was image classification. The researchers used the ImageNet database to train the system with labeled images, then to test its ability to classify novel ones. On a commercial microcontroller they tested, MCUNet successfully classified 70.7 percent of the novel images — the previous state-of-the-art neural network and inference engine combo was just 54 percent accurate. “Even a 1 percent improvement is considered significant,” says Lin. “So this is a giant leap for microcontroller settings.”
The team found similar results in ImageNet tests of three other microcontrollers. And on both speed and accuracy, MCUNet beat the competition for audio and visual “wake-word” tasks, where a user initiates an interaction with a computer using vocal cues (think: “Hey, Siri”) or simply by entering a room. The experiments highlight MCUNet’s adaptability to numerous applications.

Play video

“Huge potential”
The promising test results give Han hope that it will become the new industry standard for microcontrollers. “It has huge potential,” he says.
The advance “extends the frontier of deep neural network design even farther into the computational domain of small energy-efficient microcontrollers,” says Kurt Keutzer, a computer scientist at the University of California at Berkeley, who was not involved in the work. He adds that MCUNet could “bring intelligent computer-vision capabilities to even the simplest kitchen appliances, or enable more intelligent motion sensors.”
MCUNet could also make IoT devices more secure. “A key advantage is preserving privacy,” says Han. “You don’t need to transmit the data to the cloud.”
Analyzing data locally reduces the risk of personal information being stolen — including personal health data. Han envisions smart watches with MCUNet that don’t just sense users’ heartbeat, blood pressure, and oxygen levels, but also analyze and help them understand that information. MCUNet could also bring deep learning to IoT devices in vehicles and rural areas with limited internet access.
Plus, MCUNet’s slim computing footprint translates into a slim carbon footprint. “Our big dream is for green AI,” says Han, adding that training a large neural network can burn carbon equivalent to the lifetime emissions of five cars. MCUNet on a microcontroller would require a small fraction of that energy. “Our end goal is to enable efficient, tiny AI with less computational resources, less human resources, and less data,” says Han. More

In August, the young artificial intelligence process automation company Intelenz, Inc. announced its first U.S. patent, an AI-enabled software-as-a-service application for automating repetitive activities, improving process execution, and reducing operating costs. For company co-founder Renzo Zagni, the patent is a powerful testament to the value of his MIT educational experience.
Over the course of his two-decade career at Oracle, Zagni worked his way from database administrator to vice president of Enterprise Applications-IT. After spending seven years in his final role, he was ready to take on a new challenge by starting his own company.
From employee to entrepreneur
Zagni launched Intelenz in 2017 with a goal of keeping his company on the cutting edge. Doing so required that he stay up to date on the latest machine learning knowledge and techniques. At first, that meant exploring new concepts on his own. But to get to the next level, he realized he needed a little more formal education. That’s when he turned to MIT.
“When I discovered that I could take courses at MIT, I thought, ‘What better place to learn about artificial intelligence and machine learning?’” he says. “Access to MIT faculty was something that I simply couldn’t pass up.”
Zagni enrolled in MIT Professional Education’s Professional Certificate Program in Machine Learning and Artificial Intelligence, traveling from California to Cambridge, Massachusetts, to attend accelerated courses on the MIT campus.
As he continued to build his startup, one key to demystifying machine learning came from MIT Professor Regina Barzilay, a Delta Electronics professor in the Department of Electrical Engineering and Computer Science and a member of MIT’s Computer Science and Artificial Intelligence Laboratory. “Professor Barzilay used real-life examples in a way that helped us quickly understand very complex concepts behind machine learning and AI,” Zagni says. “And her passion and vision to use the power of machine learning to help win the fight against cancer was commendable and inspired us all.”  
The insights Zagni gained from Barzilay and other machine learning/AI faculty members helped him shape Intelenz’ early products — and continue to influence his company’s product development today — most recently, in his patented technology, the “Service Tickets Early Warning System.” The technology is an important representation of Intelenz’ ability to develop AI models aimed at automating and improving business processes at the enterprise level.
“We had a problem we wanted to solve and knew that artificial intelligence and machine learning could possibly address it. And MIT gave me the tools and the methodologies to translate these needs into a machine learning model that ended up becoming a patent,” Zagni says.
Driving machine learning with innovation
As an entrepreneur looking to push the boundaries of information technology, Zagni wasn’t content to simply use existing solutions; innovation became a key goal very early in the process.
“For professionals like me who work in information technology, innovation and artificial intelligence go hand-in-hand,” Zagni says.
While completing machine learning courses at MIT, Zagni simultaneously enrolled in MIT Professional Education’s Professional Certificate Program in Innovation and Technology. Combining his new AI knowledge with the latest approaches in innovation was a game-changer.
“During my first year with MIT, I was putting together the Intelenz team, hiring developers, and completing designs. What I learned in the innovation courses helped us a lot,” Zagni says. “For instance, Blake Kotelly‘s Mastering Innovation and Design Thinking course made a huge difference in how we develop our solutions and engage our customers. And our customers love the design-thinking approach.”
Looking forward
While his progress at Intelenz is exciting, Zagni is anything but done. As he continues to develop his organization and its AI-enabled offerings, he’s looking ahead to additional opportunities for growth.  
“We’re already looking for the next technology that is going to allow us to disrupt the market,” Zagni says. “We’re hearing a lot about quantum computing and other technology innovations. It’s very important for us to stay on top of them if we want to remain competitive.”
He remains committed to lifelong learning, and says he will definitely be looking to future MIT courses — and he recommends other professionals in his field do the same.
“Being part of the MIT ecosystem has really put me ahead of the curve by providing access to the latest information, tools, and methodologies,” Zagni says. “And on top of that, the faculty are very helpful and truly want to see participants succeed.” More

Machine learning model predicts probability that a particular urinary tract infection can be treated by specific antibiotics.  More

Dealing with a global pandemic has taken a toll on the mental health of millions of people. A team of MIT and Harvard University researchers has shown that they can measure those effects by analyzing the language that people use to express their anxiety online.
Using machine learning to analyze the text of more than 800,000 Reddit posts, the researchers were able to identify changes in the tone and content of language that people used as the first wave of the Covid-19 pandemic progressed, from January to April of 2020. Their analysis revealed several key changes in conversations about mental health, including an overall increase in discussion about anxiety and suicide.
“We found that there were these natural clusters that emerged related to suicidality and loneliness, and the amount of posts in these clusters more than doubled during the pandemic as compared to the same months of the preceding year, which is a grave concern,” says Daniel Low, a graduate student in the Program in Speech and Hearing Bioscience and Technology at Harvard and MIT and the lead author of the study.
The analysis also revealed varying impacts on people who already suffer from different types of mental illness. The findings could help psychiatrists, or potentially moderators of the Reddit forums that were studied, to better identify and help people whose mental health is suffering, the researchers say.
“When the mental health needs of so many in our society are inadequately met, even at baseline, we wanted to bring attention to the ways that many people are suffering during this time, in order to amplify and inform the allocation of resources to support them,” says Laurie Rumker, a graduate student in the Bioinformatics and Integrative Genomics PhD Program at Harvard and one of the authors of the study.
Satrajit Ghosh, a principal research scientist at MIT’s McGovern Institute for Brain Research, is the senior author of the study, which appears in the Journal of Internet Medical Research. Other authors of the paper include Tanya Talkar, a graduate student in the Program in Speech and Hearing Bioscience and Technology at Harvard and MIT; John Torous, director of the digital psychiatry division at Beth Israel Deaconess Medical Center; and Guillermo Cecchi, a principal research staff member at the IBM Thomas J. Watson Research Center.
A wave of anxiety
The new study grew out of the MIT class 6.897/HST.956 (Machine Learning for Healthcare), in MIT’s Department of Electrical Engineering and Computer Science. Low, Rumker, and Talkar, who were all taking the course last spring, had done some previous research on using machine learning to detect mental health disorders based on how people speak and what they say. After the Covid-19 pandemic began, they decided to focus their class project on analyzing Reddit forums devoted to different types of mental illness.
“When Covid hit, we were all curious whether it was affecting certain communities more than others,” Low says. “Reddit gives us the opportunity to look at all these subreddits that are specialized support groups. It’s a really unique opportunity to see how these different communities were affected differently as the wave was happening, in real-time.”
The researchers analyzed posts from 15 subreddit groups devoted to a variety of mental illnesses, including schizophrenia, depression, and bipolar disorder. They also included a handful of groups devoted to topics not specifically related to mental health, such as personal finance, fitness, and parenting.
Using several types of natural language processing algorithms, the researchers measured the frequency of words associated with topics such as anxiety, death, isolation, and substance abuse, and grouped posts together based on similarities in the language used. These approaches allowed the researchers to identify similarities between each group’s posts after the onset of the pandemic, as well as distinctive differences between groups.
The researchers found that while people in most of the support groups began posting about Covid-19 in March, the group devoted to health anxiety started much earlier, in January. However, as the pandemic progressed, the other mental health groups began to closely resemble the health anxiety group, in terms of the language that was most often used. At the same time, the group devoted to personal finance showed the most negative semantic change from January to April 2020, and significantly increased the use of words related to economic stress and negative sentiment.
They also discovered that the mental health groups affected the most negatively early in the pandemic were those related to ADHD and eating disorders. The researchers hypothesize that without their usual social support systems in place, due to lockdowns, people suffering from those disorders found it much more difficult to manage their conditions. In those groups, the researchers found posts about hyperfocusing on the news and relapsing back into anorexia-type behaviors since meals were not being monitored by others due to quarantine.
Using another algorithm, the researchers grouped posts into clusters such as loneliness or substance use, and then tracked how those groups changed as the pandemic progressed. Posts related to suicide more than doubled from pre-pandemic levels, and the groups that became significantly associated with the suicidality cluster during the pandemic were the support groups for borderline personality disorder and post-traumatic stress disorder.
The researchers also found the introduction of new topics specifically seeking mental health help or social interaction. “The topics within these subreddit support groups were shifting a bit, as people were trying to adapt to a new life and focus on how they can go about getting more help if needed,” Talkar says.
While the authors emphasize that they cannot implicate the pandemic as the sole cause of the observed linguistic changes, they note that there was much more significant change during the period from January to April in 2020 than in the same months in 2019 and 2018, indicating the changes cannot be explained by normal annual trends.
Mental health resources
This type of analysis could help mental health care providers identify segments of the population that are most vulnerable to declines in mental health caused by not only the Covid-19 pandemic but other mental health stressors such as controversial elections or natural disasters, the researchers say.
Additionally, if applied to Reddit or other social media posts in real-time, this analysis could be used to offer users additional resources, such as guidance to a different support group, information on how to find mental health treatment, or the number for a suicide hotline.
“Reddit is a very valuable source of support for a lot of people who are suffering from mental health challenges, many of whom may not have formal access to other kinds of mental health support, so there are implications of this work for ways that support within Reddit could be provided,” Rumker says.
The researchers now plan to apply this approach to study whether posts on Reddit and other social media sites can be used to detect mental health disorders. One current project involves screening posts in a social media site for veterans for suicide risk and post-traumatic stress disorder.
The research was funded by the National Institutes of Health and the McGovern Institute. More

GPS isn’t waterproof. The navigation system depends on radio waves, which break down rapidly in liquids, including seawater. To track undersea objects like drones or whales, researchers rely on acoustic signaling. But devices that generate and send sound usually require batteries — bulky, short-lived batteries that need regular changing. Could we do without them?
MIT researchers think so. They’ve built a battery-free pinpointing system dubbed Underwater Backscatter Localization (UBL). Rather than emitting its own acoustic signals, UBL reflects modulated signals from its environment. That provides researchers with positioning information, at net-zero energy. Though the technology is still developing, UBL could someday become a key tool for marine conservationists, climate scientists, and the U.S. Navy.
These advances are described in a paper being presented this week at the Association for Computing Machinery’s Hot Topics in Networks workshop, by members of the Media Lab’s Signal Kinetics group. Research Scientist Reza Ghaffarivardavagh led the paper, along with co-authors Sayed Saad Afzal, Osvy Rodriguez, and Fadel Adib, who leads the group and is the Doherty Chair of Ocean Utilization as well as an associate professor in the MIT Media Lab and the MIT Department of Electrical Engineering and Computer Science.
“Power-hungry”
It’s nearly impossible to escape GPS’ grasp on modern life. The technology, which relies on satellite-transmitted radio signals, is used in shipping, navigation, targeted advertising, and more. Since its introduction in the 1970s and ’80s, GPS has changed the world. But it hasn’t changed the ocean. If you had to hide from GPS, your best bet would be underwater.
Because radio waves quickly deteriorate as they move through water, subsea communications often depend on acoustic signals instead. Sound waves travel faster and further underwater than through air, making them an efficient way to send data. But there’s a drawback.
“Sound is power-hungry,” says Adib. For tracking devices that produce acoustic signals, “their batteries can drain very quickly.” That makes it hard to precisely track objects or animals for a long time-span — changing a battery is no simple task when it’s attached to a migrating whale. So, the team sought a battery-free way to use sound.
Good vibrations
Adib’s group turned to a unique resource they’d previously used for low-power acoustic signaling: piezoelectric materials. These materials generate their own electric charge in response to mechanical stress, like getting pinged by vibrating soundwaves. Piezoelectric sensors can then use that charge to selectively reflect some soundwaves back into their environment. A receiver translates that sequence of reflections, called backscatter, into a pattern of 1s (for soundwaves reflected) and 0s (for soundwaves not reflected). The resulting binary code can carry information about ocean temperature or salinity.
In principle, the same technology could provide location information. An observation unit could emit a soundwave, then clock how long it takes that soundwave to reflect off the piezoelectric sensor and return to the observation unit. The elapsed time could be used to calculate the distance between the observer and the piezoelectric sensor. But in practice, timing such backscatter is complicated, because the ocean can be an echo chamber.
The sound waves don’t just travel directly between the observation unit and sensor. They also careen between the surface and seabed, returning to the unit at different times. “You start running into all of these reflections,” says Adib. “That makes it complicated to compute the location.” Accounting for reflections is an even greater challenge in shallow water — the short distance between seabed and surface means the confounding rebound signals are stronger.
The researchers overcame the reflection issue with “frequency hopping.” Rather than sending acoustic signals at a single frequency, the observation unit sends a sequence of signals across a range of frequencies. Each frequency has a different wavelength, so the reflected sound waves return to the observation unit at different phases. By combining information about timing and phase, the observer can pinpoint the distance to the tracking device. Frequency hopping was successful in the researchers’ deep-water simulations, but they needed an additional safeguard to cut through the reverberating noise of shallow water.
Where echoes run rampant between the surface and seabed, the researchers had to slow the flow of information. They reduced the bitrate, essentially waiting longer between each signal sent out by the observation unit. That allowed the echoes of each bit to die down before potentially interfering with the next bit. Whereas a bitrate of 2,000 bits/second sufficed in simulations of deep water, the researchers had to dial it down to 100 bits/second in shallow water to obtain a clear signal reflection from the tracker. But a slow bitrate didn’t solve everything.
To track moving objects, the researchers actually had to boost the bitrate. One thousand bits/second was too slow to pinpoint a simulated object moving through deep water at 30 centimeters/second. “By the time you get enough information to localize the object, it has already moved from its position,” explains Afzal. At a speedy 10,000 bits/second, they were able to track the object through deep water.
Efficient exploration
Adib’s team is working to improve the UBL technology, in part by solving challenges like the conflict between low bitrate required in shallow water and the high bitrate needed to track movement. They’re working out the kinks through tests in the Charles River. “We did most of the experiments last winter,” says Rodriguez. That included some days with ice on the river. “It was not very pleasant.”
Conditions aside, the tests provided a proof-of-concept in a challenging shallow-water environment. UBL estimated the distance between a transmitter and backscatter node at various distances up to nearly half a meter. The team is working to increase UBL’s range in the field, and they hope to test the system with their collaborators at the Wood Hole Oceanographic Institution on Cape Cod.
They hope UBL can help fuel a boom in ocean exploration. Ghaffarivardavagh notes that scientists have better maps of the moon’s surface than of the ocean floor. “Why can’t we send out unmanned underwater vehicles on a mission to explore the ocean? The answer is: We will lose them,” he says.
UBL could one day help autonomous vehicles stay found underwater, without spending precious battery power. The technology could also help subsea robots work more precisely, and provide information about climate change impacts in the ocean. “There are so many applications,” says Adib. “We’re hoping to understand the ocean at scale. It’s a long-term vision, but that’s what we’re working toward and what we’re excited about.”
This work was supported, in part, by the Office of Naval Research. More

Asymptomatic people who are infected with Covid-19 exhibit, by definition, no discernible physical symptoms of the disease. They are thus less likely to seek out testing for the virus, and could unknowingly spread the infection to others.
But it seems those who are asymptomatic may not be entirely free of changes wrought by the virus. MIT researchers have now found that people who are asymptomatic may differ from healthy individuals in the way that they cough. These differences are not decipherable to the human ear. But it turns out that they can be picked up by artificial intelligence.
In a paper published recently in the IEEE Journal of Engineering in Medicine and Biology, the team reports on an AI model that distinguishes asymptomatic people from healthy individuals through forced-cough recordings, which people voluntarily submitted through web browsers and devices such as cellphones and laptops.
The researchers trained the model on tens of thousands of samples of coughs, as well as spoken words. When they fed the model new cough recordings, it accurately identified 98.5 percent of coughs from people who were confirmed to have Covid-19, including 100 percent of coughs from asymptomatics — who reported they did not have symptoms but had tested positive for the virus.
The team is working on incorporating the model into a user-friendly app, which if FDA-approved and adopted on a large scale could potentially be a free, convenient, noninvasive prescreening tool to identify people who are likely to be asymptomatic for Covid-19. A user could log in daily, cough into their phone, and instantly get information on whether they might be infected and therefore should confirm with a formal test.
“The effective implementation of this group diagnostic tool could diminish the spread of the pandemic if everyone uses it before going to a classroom, a factory, or a restaurant,” says co-author Brian Subirana, a research scientist in MIT’s Auto-ID Laboratory.
Subirana’s co-authors are Jordi Laguarta and Ferran Hueto, of MIT’s Auto-ID Laboratory.

Vocal sentiments
Prior to the pandemic’s onset, research groups already had been training algorithms on cellphone recordings of coughs to accurately diagnose conditions such as pneumonia and asthma. In similar fashion, the MIT team was developing AI models to analyze forced-cough recordings to see if they could detect signs of Alzheimer’s, a disease associated with not only memory decline but also neuromuscular degradation such as weakened vocal cords.
They first trained a general machine-learning algorithm, or neural network, known as ResNet50, to discriminate sounds associated with different degrees of vocal cord strength. Studies have shown that the quality of the sound “mmmm” can be an indication of how weak or strong a person’s vocal cords are. Subirana trained the neural network on an audiobook dataset with more than 1,000 hours of speech, to pick out the word “them” from other words like “the” and “then.”
The team trained a second neural network to distinguish emotional states evident in speech, because Alzheimer’s patients — and people with neurological decline more generally — have been shown to display certain sentiments such as frustration, or having a flat affect, more frequently than they express happiness or calm. The researchers developed a sentiment speech classifier model by training it on a large dataset of actors intonating emotional states, such as neutral, calm, happy, and sad.
The researchers then trained a third neural network on a database of coughs in order to discern changes in lung and respiratory performance.
Finally, the team combined all three models, and overlaid an algorithm to detect muscular degradation. The algorithm does so by essentially simulating an audio mask, or layer of noise, and distinguishing strong coughs — those that can be heard over the noise — over weaker ones.
With their new AI framework, the team fed in audio recordings, including of Alzheimer’s patients, and found it could identify the Alzheimer’s samples better than existing models. The results showed that, together, vocal cord strength, sentiment, lung and respiratory performance, and muscular degradation were effective biomarkers for diagnosing the disease.
When the coronavirus pandemic began to unfold, Subirana wondered whether their AI framework for Alzheimer’s might also work for diagnosing Covid-19, as there was growing evidence that infected patients experienced some similar neurological symptoms such as temporary neuromuscular impairment.
“The sounds of talking and coughing are both influenced by the vocal cords and surrounding organs. This means that when you talk, part of your talking is like coughing, and vice versa. It also means that things we easily derive from fluent speech, AI can pick up simply from coughs, including things like the person’s gender, mother tongue, or even emotional state. There’s in fact sentiment embedded in how you cough,” Subirana says. “So we thought, why don’t we try these Alzheimer’s biomarkers [to see if they’re relevant] for Covid.”
“A striking similarity”
In April, the team set out to collect as many recordings of coughs as they could, including those from Covid-19 patients. They established a website where people can record a series of coughs, through a cellphone or other web-enabled device. Participants also fill out a survey of symptoms they are experiencing, whether or not they have Covid-19, and whether they were diagnosed through an official test, by a doctor’s assessment of their symptoms, or if they self-diagnosed. They also can note their gender, geographical location, and native language.
To date, the researchers have collected more than 70,000 recordings, each containing several coughs, amounting to some 200,000 forced-cough audio samples, which Subirana says is “the largest research cough dataset that we know of.” Around 2,500 recordings were submitted by people who were confirmed to have Covid-19, including those who were asymptomatic.
The team used the 2,500 Covid-associated recordings, along with 2,500 more recordings that they randomly selected from the collection to balance the dataset. They used 4,000 of these samples to train the AI model. The remaining 1,000 recordings were then fed into the model to see if it could accurately discern coughs from Covid patients versus healthy individuals.
Surprisingly, as the researchers write in their paper, their efforts have revealed “a striking similarity between Alzheimer’s and Covid discrimination.”
Without much tweaking within the AI framework originally meant for Alzheimer’s, they found it was able to pick up patterns in the four biomarkers — vocal cord strength, sentiment, lung and respiratory performance, and muscular degradation — that are specific to Covid-19. The model identified 98.5 percent of coughs from people confirmed with Covid-19, and of those, it accurately detected all of the asymptomatic coughs.
“We think this shows that the way you produce sound, changes when you have Covid, even if you’re asymptomatic,” Subirana says.
Asymptomatic symptoms
The AI model, Subirana stresses, is not meant to diagnose symptomatic people, as far as whether their symptoms are due to Covid-19 or other conditions like flu or asthma. The tool’s strength lies in its ability to discern asymptomatic coughs from healthy coughs.  
The team is working with a company to develop a free pre-screening app based on their AI model. They are also partnerning with several hospitals around the world to collect a larger, more diverse set of cough recordings, which will help to train and strengthen the model’s accuracy.
As they propose in their paper, “Pandemics could be a thing of the past if pre-screening tools are always on in the background and constantly improved.”
Ultimately, they envision that audio AI models like the one they’ve developed may be incorporated into smart speakers and other listening devices so that people can conveniently get an initial assessment of their disease risk, perhaps on a daily basis.
This research was supported, in part, by Takeda Pharmaceutical Company Limited. More