Live imaging method brings structure to mapping brain function

To understand the massive capabilities and complexities of the brain, neuroscientists segment it into regions based on what they appear to do — such as processing what we sense, or how to move. What’s been lacking, however, is an ability to tie those functional maps precisely and consistently to matching distinctions of physical structure, especially in live animals while they are performing the functions of interest. In a new study, MIT researchers demonstrate a new way to do that, providing an unprecedented pairing of functional mapping in live mice with distinguishing structural information for each region all the way through the cortex into deeper tissue below.

“Our study shows for the first time that structural and functional coupling of visual areas in the mouse brain can be detected at sub-cellular resolution in vivo,” wrote the authors based in the lab of Mriganka Sur, the Newton Professor of Neuroscience in The Picower Institute for Learning and Memory and the Department of Brain and Cognitive Sciences at MIT.

The technique could give scientists more precise ways to distinguish the borders and contents of regions they wish to study and could help them better understand the way that structural distinctions develop within individuals in different functional regions over time. Sur’s lab, for instance, is intensely interested in understanding the especially complex development of vision. Humans have 35 or so distinct functional regions that contribute to processing vision, Sur notes, and even mice have 10.

“There is something profound in the way that vision is represented and created in mammalian brains,” Sur says. “Where do these areas come from, what do they mean and what do they do? It has not been easy to understand how they differ. The critical thing is to precisely map or match the functional representation of each area with its anatomical uniqueness.”

To develop tools to help answer those questions, postdoc Murat Yildirim led the study published in Biomedical Optics Express. In it he describes how the research team combined a method of charting functional areas — retinotopic mapping — with deep structural information measured by a technology he has helped to pioneer — third-harmonic generation (THG) three-photon microscopy.

In retinotopic mapping, researchers can identify functional regions by engineering neurons to flash when they become electrically active (and show changes in calcium) in response to a particular stimulation. For example, scientists could show a mouse a pattern moving across a screen and mark where neurons light up, with each area showing a characteristic location and pattern of response.

Three-photon microscopy can finely resolve individual cells and their smaller substructures as deep as a millimeter or more — enough to see all the way through the cortex. THG, meanwhile, adds the capability to finely resolve both blood vessels and the fibers of a material called myelin that wrap the long, tendrilous axons of many neurons. THG does not require adding any labeling dyes or chemicals.

Crucially, THG yields an important optical measure called effective attenuation length (EAL), which is a measure of how much the light is absorbed or scattered as it moves through the tissue. In the study, Yildirim and co-authors show that EAL specifically depends on each region’s unique architecture of cells, blood vessels, and myelin. They measured EAL in each of six visual functional regions and showed that the EAL significantly differed among neighboring visual areas, providing a structural signature of sorts for each functional area. Their measurements were so precise, in fact, that they could show how EAL varied within functional regions, being most unique toward the middle and blending closer to the values of neighboring regions out toward the borders.

In other words, by combining the retinotopic mapping with THG three-photon microscopy, Yildirim said, scientists can identify distinct regions by both their function and structure while continuing to work with animals in live experiments. This can produce more accurate and faster results than making observations during behavior and then dissecting tissue in hopes of relocating those same exact positions in preserved brain sections later.

“We would like to combine the strength of retinotopic mapping with three-photon imaging to get more structural information,” Yildirim says. “Otherwise there may be some discrepancies when you do the live imaging of brain activity but then take the tissue out, stain it and try to find the same region.”

Especially as three-photon microscopy gains wider adoption and imaging speeds improve — right now, imaging a millimeter-deep column of cortex takes about 15 minutes, the authors acknowledge — the team expects that its new method could be used not only for studies of the visual system, but also in regions all around the cortex. Moreover, it may help characterize disease states as well as healthy brain structure and function.

“This advance should enable similar studies of structural and functional coupling in other sensory and non-sensory cortical areas in the brains of mice and other animal models,” they wrote. “We believe that the structural and functional correlation in visual areas that we describe for the first time points to crucial developmental mechanisms that set up these areas, thus our work would lead to a better fundamental understanding of brain development, and of disorders such as Alzheimer’s, stroke, and aging.”

In addition to Yildirim and Sur, the paper’s other authors are Ming Hu, Nhat Le, Hiroki Sugihara, and Peter So.

The National Institutes of Health, the National Science Foundation, The JPB Foundation and the Massachusetts Life Sciences Initiative provided funding for the study.

de MIT News https://ift.tt/3i1JY4O

Provably exact artificial intelligence for nuclear and particle physics

The Standard Model of particle physics describes all the known elementary particles and three of the four fundamental forces governing the universe; everything except gravity. These three forces — electromagnetic, strong, and weak — govern how particles are formed, how they interact, and how the particles decay.

Studying particle and nuclear physics within this framework, however, is difficult, and relies on large-scale numerical studies. For example, many aspects of the strong force require numerically simulating the dynamics at the scale of 1/10th to 1/100th the size of a proton to answer fundamental questions about the properties of protons, neutrons, and nuclei.

“Ultimately, we are computationally limited in the study of proton and nuclear structure using lattice field theory,” says assistant professor of physics Phiala Shanahan. “There are a lot of interesting problems that we know how to address in principle, but we just don’t have enough compute, even though we run on the largest supercomputers in the world.”

To push past these limitations, Shanahan leads a group that combines theoretical physics with machine learning models. In their paper “Equivariant flow-based sampling for lattice gauge theory,” published this month in Physical Review Letters, they show how incorporating the symmetries of physics theories into machine learning and artificial intelligence architectures can provide much faster algorithms for theoretical physics. 

“We are using machine learning not to analyze large amounts of data, but to accelerate first-principles theory in a way which doesn’t compromise the rigor of the approach,” Shanahan says. “This particular work demonstrated that we can build machine learning architectures with some of the symmetries of the Standard Model of particle and nuclear physics built in, and accelerate the sampling problem we are targeting by orders of magnitude.” 

Shanahan launched the project with MIT graduate student Gurtej Kanwar and with Michael Albergo, who is now at NYU. The project expanded to include Center for Theoretical Physics postdocs Daniel Hackett and Denis Boyda, NYU Professor Kyle Cranmer, and physics-savvy machine-learning scientists at Google Deep Mind, Sébastien Racanière and Danilo Jimenez Rezende.

This month’s paper is one in a series aimed at enabling studies in theoretical physics that are currently computationally intractable. “Our aim is to develop new algorithms for a key component of numerical calculations in theoretical physics,” says Kanwar. “These calculations inform us about the inner workings of the Standard Model of particle physics, our most fundamental theory of matter. Such calculations are of vital importance to compare against results from particle physics experiments, such as the Large Hadron Collider at CERN, both to constrain the model more precisely and to discover where the model breaks down and must be extended to something even more fundamental.”

The only known systematically controllable method of studying the Standard Model of particle physics in the nonperturbative regime is based on a sampling of snapshots of quantum fluctuations in the vacuum. By measuring properties of these fluctuations, once can infer properties of the particles and collisions of interest.

This technique comes with challenges, Kanwar explains. “This sampling is expensive, and we are looking to use physics-inspired machine learning techniques to draw samples far more efficiently,” he says. “Machine learning has already made great strides on generating images, including, for example, recent work by NVIDIA to generate images of faces 'dreamed up' by neural networks. Thinking of these snapshots of the vacuum as images, we think it's quite natural to turn to similar methods for our problem.”

Adds Shanahan, “In our approach to sampling these quantum snapshots, we optimize a model that takes us from a space that is easy to sample to the target space: given a trained model, sampling is then efficient since you just need to take independent samples in the easy-to-sample space, and transform them via the learned model.”

In particular, the group has introduced a framework for building machine-learning models that exactly respect a class of symmetries, called "gauge symmetries," crucial for studying high-energy physics.

As a proof of principle, Shanahan and colleagues used their framework to train machine-learning models to simulate a theory in two dimensions, resulting in orders-of-magnitude efficiency gains over state-of-the-art techniques and more precise predictions from the theory. This paves the way for significantly accelerated research into the fundamental forces of nature using physics-informed machine learning.

The group’s first few papers as a collaboration discussed applying the machine-learning technique to a simple lattice field theory, and developed this class of approaches on compact, connected manifolds which describe the more complicated field theories of the Standard Model. Now they are working to scale the techniques to state-of-the-art calculations.

“I think we have shown over the past year that there is a lot of promise in combining physics knowledge with machine learning techniques,” says Kanwar. “We are actively thinking about how to tackle the remaining barriers in the way of performing full-scale simulations using our approach. I hope to see the first application of these methods to calculations at scale in the next couple of years. If we are able to overcome the last few obstacles, this promises to extend what we can do with limited resources, and I dream of performing calculations soon that give us novel insights into what lies beyond our best understanding of physics today.”

This idea of physics-informed machine learning is also known by the team as “ab-initio AI,” a key theme of the recently launched MIT-based National Science Foundation Institute for Artificial Intelligence and Fundamental Interactions (IAIFI), where Shanahan is research coordinator for physics theory.

Led by the Laboratory for Nuclear Science, the IAIFI is comprised of both physics and AI researchers at MIT and Harvard, Northeastern, and Tufts universities.

“Our collaboration is a great example of the spirit of IAIFI, with a team with diverse backgrounds coming together to advance AI and physics simultaneously” says Shanahan. As well as research like Shanahan’s targeting physics theory, IAIFI researchers are also working to use AI to enhance the scientific potential of various facilities, including the Large Hadron Collider and the Laser Interferometer Gravity Wave Observatory, and to advance AI itself. 

de MIT News https://ift.tt/3hY9urG

How MIT’s rocket, electric car, and solar vehicle teams have adjusted to the pandemic

The Covid-19 pandemic has hit all areas of the MIT communty. Those for which in-person hands-on experiences traditionally take center stage have had to pivot in unique ways.

In this Q&A, the MIT Edgerton Center checked in with with leaders of three of the Institute's student teams that traditionally plan their activities around engineering vehicles or devices in preparation for national or international competitions. Here, juniors Max Kwon, president of the MIT Rocket Team; Jorge Nin, MIT Motorsports captain; and Aditya Mehrotra, captain of MIT Solar Electric Vehicle Team describe how the pandemic has affected their operations and plans for the future. Interviews were edited for length and clarity.

Q: Could you introduce your teams to the MIT community?

Max Kwon: The MIT Rocket team is comprised of a majority of undergraduate students, with a few graduate students. Our aim is to build a two-stage rocket that crosses the Kármán Line, which is the recognized boundary of space.

Jorge Nin: MIT Motorsports is a student team for undergrads who design, manufacture, build, and test a formula-style electric race car. At the end of the year, we end up competing against other [student] teams.

Aditya Mehrota: MIT Solar Team was founded 30 years ago, so it’s one of the oldest teams on campus. We design and build solar powered vehicles to race in international solar competitions.

Q: All three of you recently assumed leadership of your teams. What did your initial plans look like?

Kwon: When I became president in January, I had a specific vision for what the team would be doing for the year. We were working on Project Phoenix, a two-stage rocket. By the end of spring 2020, we wanted to independently launch our second stage, just as a test. [So the plan was that] we build the second stage, we take it out, probably some rocket launch field in New York, and we test our system there. It’s always fun to go out and see a rocket launch. And then over the summer and the fall, we keep developing this rocket, we come back, we build the first stage, and then in IAP of 2021, we’d go out to California into the Mojave Desert and launch the full two-stage rocket.

Nin: So right before Covid-19 hit, we had just hit one of our major milestones. There are three big ones. The first is the design phase. After that we have “rolling,” when the car has the wheels on it and we can push it forward, and you can turn the steering wheel. And then you have “Roman,” the first time that the car drives on its own power. We had just hit “Roman.” In fact, it was the fastest we've ever hit “Roman” in recent team memory.

Mehrotra: Our team operates on a two-year cycle: our first year is a design year, and the second is a build year. This year we were building [the car]. On spring break, we were going to take the car on a road trip to test it. And what this would have allowed us to do was spend the rest of that semester fixing up the car and getting the team ready for the American Solar Challenge 2020.

Q: And then of course, the pandemic happened. In March, MIT asked students to leave campus. What did that mean for your team’s plans?

Kwon: It was all thrown up in the air. Obviously, we couldn't launch the second stage. We couldn't have that test launch at the end of the semester.

Nin: We’d just hit “Roman.” And we were really excited because we were on a really good path forward to win the summer competition. And then we were informed that we would all have to leave campus. It was very sad — a week later we were going to go to Palmer racetrack to try out the car for the first time.

Mehrotra: We were supposed to be racing the vehicle this summer. But then students were asked to leave campus because of the pandemic. We weren't able to go out and test the car. And of course, we were unable to race this summer. The good news was, because our car drives on real roads, we needed to get the car registered with the DMV [and we] were able to get the car registered the last week before students left for campus. So we should be getting our license plate soon. So that was kind of a pick-me-up for people.

Q: How did your teams confront the crisis?

Kwon: It definitely was difficult. But I think the entire team has done a really good job of staying on top of things. I'd honestly say we were pretty lucky that we weren't in a manufacturing phase. We were in a design phase. So we knew that there [was] stuff we could do. What quarantine has allowed us to do is slow down, take a step back, and kind of run through the engineering design process. The way it should be.

Nin: We had all these crazy ideas. Like maybe we ship this car to Los Angeles, and then we test over there for a little bit. In the end, we ended up leaving the car in the garage and we all went home. And that's when we decided to transfer over executive power from last year's team. So, what we started doing after that is planning out what we would do for the next year and how we could take advantage of the situation.

We also had a unique challenge in that this year we didn't have many sophomores. We had a lot of seniors, a lot of first-years. That was a big issue in terms of our knowledge transfer because we're going to have a lot of information that was going to be lost if you couldn't transfer it over to the members. And that's really what we're trying to focus on during this time: knowledge transfer.

Mehrotra: Of course, the team was disappointed. But at the moment our team is laying low a little bit because there's not that much we can really do right now. To finish the car, we need to be [there] in person. The race wasn't fully canceled, it was postponed to next summer. So [American Solar Challenge] is going to happen in the summer of 2021. So our team is going to start focusing on preparing for that as soon as we know the plan going forward.

Q: How did MIT's announcement that only seniors would return to campus this semester change things for your team?

Kwon: The original plan was to come back in the fall and begin work on the manufacturing of the second stage of the rocket and then launch that second stage during IAP [Independent Activities Period, in January]. Because only seniors were allowed to go back to campus, [this was] not a possibility. This semester, we will have some seniors who will be able to do small manufacturing tasks. But we are going to be nowhere near ready to launch by [January]. Our stated goal is still to launch the full stack — the first and the second stage, together — in the summer of 2021. But that's a goal that will also have to be reassessed depending on what the spring semester looks like. But because we're not doing so much with the rocket over the fall, we're diverting a lot of our energy into recruitment and team skills building. 

Nin: The biggest change really has just been focusing more over the fall on R&D projects and trying to make sure that we can transfer as much knowledge from the seniors to the rest of the members. In spring, we will focus more on getting the car up to speed, because we will have limited people and we won't have an IAP to build a new car. It just becomes very difficult to build one from scratch.

Mehrotra: It does change our timeline a little bit. What we're probably going to end up doing is going to be more focused on things like design, integrating new members into the team. And I think we're holding off until the spring to really decide what we want to do because by January, I think we'll have a more accurate understanding based on how the situation progresses on whether we're going to be able to compete this summer or if we're going to have to hold off. And maybe just race the car we're currently building not this summer, but two summers from now.

Q: What would you like the MIT community to know about your teams? 

Kwon: I'm sure it's the same for all Edgerton Center teams. We've all been hit the same way. And I'm just personally impressed with the team and how we've been able to go through quarantine and be productive. [The team has] been super supportive of helping us get through quarantine and meet our goals that we've had for the semester. And I think the fact that our rocket team has pushed so, so hard, even through quarantine — it really just shows the spirit of MIT.

To experience building something with your peers and watching it, literally launch and fly, is such a rewarding experience that I hope everyone outside the Edgerton Center can appreciate and see. The work that the EC is doing and the work that the teams are doing inside of it really is something that matters and is what makes MIT, MIT.

Nin: I really, really love how much people put effort into trying to build a car and how much it means for people and the great friendships that are formed. Both through the good times and especially through the bad times when a project isn't working at 2 a.m. and you just have to pull through to get it ready for a competition.

Mehrotra: Having a good team around you takes a lot of stress off of you when making decisions. So I'm fortunate enough to have good team members not only on the regular team but on my executives as well. I really love our team because of the focus our team puts on learning and the opportunities students get from joining teams like ours. And more importantly than just for MIT students, I think our team has always been a team that tries to do like community outreach to whatever extent we can.

de MIT News https://ift.tt/3kQAg7f

Why social media has changed the world — and how to fix it

Are you on social media a lot? When is the last time you checked Twitter, Facebook, or Instagram? Last night? Before breakfast? Five minutes ago?

If so, you are not alone — which is the point, of course. Humans are highly social creatures. Our brains have become wired to process social information, and we usually feel better when we are connected. Social media taps into this tendency.

“Human brains have essentially evolved because of sociality more than any other thing,” says Sinan Aral, an MIT professor and expert in information technology and marketing. “When you develop a population-scale technology that delivers social signals to the tune of trillions per day in real-time, the rise of social media isn’t unexpected. It’s like tossing a lit match into a pool of gasoline.”

The numbers make this clear. In 2005, about 7 percent of American adults used social media. But by 2017, 80 percent of American adults used Facebook alone. About 3.5 billion people on the planet, out of 7.7 billion, are active social media participants. Globally, during a typical day, people post 500 million tweets, share over 10 billion pieces of Facebook content, and watch over a billion hours of YouTube video.

As social media platforms have grown, though, the once-prevalent, gauzy utopian vision of online community has disappeared. Along with the benefits of easy connectivity and increased information, social media has also become a vehicle for disinformation and political attacks from beyond sovereign borders.

“Social media disrupts our elections, our economy, and our health,” says Aral, who is the David Austin Professor of Management at the MIT Sloan School of Management.

Now Aral has written a book about it. In “The Hype Machine,” published this month by Currency, a Random House imprint, Aral details why social media platforms have become so successful yet so problematic, and suggests ways to improve them.

The book covers some of the same territory as “The Social Dilemma,” a popular documentary on Netflix. But Aral’s book, as he puts it, "starts where ‘The Social Dilemma’ leaves off and goes one step further to ask: What can we do to achieve the promise of social media and avoid its peril?”

“This machine exists in every facet of our lives,” Aral says. “And the question in the book is, what do we do? How do we achieve the promise of this machine and avoid the peril? We’re at a crossroads. What we do next is essential, so I want to equip people, policymakers, and platforms to help us achieve the good outcomes and avoid the bad outcomes.”

When “engagement” equals anger

“The Hype Machine” draws on Aral’s own research about social networks, as well as other findings, from the cognitive sciences, computer science, business, politics, and more. Researchers at the University of California at Los Angeles, for instance, have found that people obtain bigger hits of dopamine — the chemical in our brains highly bound up with motivation and reward — when their social media posts receive more likes.

At the same time, consider a 2018 MIT study by Soroush Vosoughi, an MIT PhD student and now an assistant professor of computer science at Dartmouth College; Deb Roy, MIT professor of media arts and sciences and executive director of the MIT Media Lab; and Aral, who has been studying social networking for 20 years. The three researchers found that on Twitter, from 2006 to 2017, false news stories were 70 percent more likely to be retweeted than true ones. Why? Most likely because false news has greater novelty value compared to the truth, and provokes stronger reactions — especially disgust and surprise.

In this light, the essential tension surrounding social media companies is that their platforms gain audiences and revenue when posts provoke strong emotional responses, often based on dubious content.

“This is a well-designed, well-thought-out machine that has objectives it maximizes,” Aral says. “The business models that run the social-media industrial complex have a lot to do with the outcomes we’re seeing — it’s an attention economy, and businesses want you engaged. How do they get engagement? Well, they give you little dopamine hits, and … get you riled up. That’s why I call it the hype machine. We know strong emotions get us engaged, so [that favors] anger and salacious content.”

From Russia to marketing

“The Hype Machine” explores both the political implications and business dimensions of social media in depth. Certainly social media is fertile terrain for misinformation campaigns. During the 2016 U.S. presidential election, Russia spread  false information to at least 126 million people on Facebook and another 20 million people on Insta­gram (which Facebook owns), and was responsible for 10 million tweets. About 44 percent of adult Americans visited a false news source in the final weeks of the campaign.

“I think we need to be a lot more vigilant than we are,” says Aral.

We do not know if Russia’s efforts altered the outcome of the 2016 election, Aral says, though they may have been fairly effective. Curiously, it is not clear if the same is true of most U.S. corporate engagement efforts.

As Aral examines, digital advertising on most big U.S. online platforms is often wildly ineffective, with academic studies showing that the “lift” generated by ad campaigns — the extent to which they affect consumer action — has been overstated by a factor of hundreds, in some cases. Simply counting clicks on ads is not enough. Instead, online engagement tends to be more effective among new consumers, and when it is targeted well; in that sense, there is a parallel between good marketing and guerilla social media campaigns.

“The two questions I get asked the most these days,” Aral says, “are, one, did Russia succeed in intervening in our democracy? And two, how do I measure the ROI [return on investment] from marketing investments? As I was writing this book, I realized the answer to those two questions is the same.”

Ideas for improvement

“The Hype Machine” has received praise from many commentators. Foster Provost, a professor at New York University’s Stern School of Business, says it is a “masterful integration of science, business, law, and policy.” Duncan Watts, a university professor at the University of Pennsylvania, says the book is “essential reading for anyone who wants to understand how we got here and how we can get somewhere better.”

In that vein, “The Hype Machine” has several detailed suggestions for improving social media. Aral favors automated and user-generated labeling of false news, and limiting revenue-collection that is based on false content. He also calls for firms to help scholars better research the issue of election interference.

Aral believes federal privacy measures could be useful, if we learn from the benefits and missteps of the General Data Protection Regulation (GDPR) in Europe and a new California law that lets consumers stop some data-sharing and allows people to find out what information companies have stored about them. He does not endorse breaking up Facebook, and suggests instead that the social media economy needs structural reform. He calls for data portability and interoperability, so “consumers would own their identities and could freely switch from one network to another.” Aral believes that without such fundamental changes, new platforms will simply replace the old ones, propelled by the network effects that drive the social-media economy.

“I do not advocate any one silver bullet,” says Aral, who emphasizes that changes in four areas together — money, code, norms, and laws — can alter the trajectory of the social media industry.

But if things continue without change, Aral adds, Facebook and the other social media giants risk substantial civic backlash and user burnout.

“If you get me angry and riled up, I might click more in the short term, but I might also grow really tired and annoyed by how this is making my life miserable, and I might turn you off entirely,” Aral observes. “I mean, that’s why we have a Delete Facebook movement, that’s why we have a Stop Hate for Profit movement. People are pushing back against the short-term vision, and I think we need to embrace this longer-term vision of a healthier communications ecosystem.”

Changing the social media giants can seem like a tall order. Still, Aral says, these firms are not necessarily destined for domination.

“I don’t think this technology or any other technology has some deterministic endpoint,” Aral says. “I want to bring us back to a more practical reality, which is that technology is what we make it, and we are abdicating our responsibility to steer technology toward good and away from bad. That is the path I try to illuminate in this book.”

de MIT News https://ift.tt/2RVG8iR

National Science Foundation Convergence Accelerator awards two grants to MIT

Two grants have been awarded to MIT researchers on the themes of socio-resilient infrastructure, and on the future of oceans. The grants are part of the U.S. National Science Foundation Convergence Accelerator program, designed to foster global cross-disciplinary and cross-sector workshops on emerging areas of critical societal importance. The NSF Convergence Accelerator program further aims to accelerate use-inspired, convergence research via partnerships between academic and non-academic stakeholders.

Socio-resilient infrastructures

The Socioresilient Infrastructure: Precision Materials, Assemblages, and Systems project is co-led by Christine Ortiz, the Morris Cohen Professor of Materials Science and Engineering, and Ellan Spero, a historian of science and technology and instructor in the Department of Materials Science and Engineering. This project will engage leading researchers from around the world to advance an intellectual framework for socio-resilient infrastructure, where social resilience is considered to be the ability of human communities to cope with and adapt to stresses and shocks such as social, political, environmental, or economic change.

This workshop will integrate exciting advances across length scales from materials (i.e., materials science and engineering, chemistry, and mechanical engineering) to assemblages (i.e., civil, structural, and environmental engineering; architecture; art and design) to systems (i.e., engineering systems, computer science and engineering, urban studies and planning including civic design and engagement). It will incorporate and center cross-cutting humanistic and socially-focused research in material culture, social justice, equity-based, community and participatory co-design, environmental and social life cycle assessment, socio-technical and sociological analysis (i.e., social sciences, STS, history).

Broad topic areas will include emerging approaches to socio-resilient and circular materials design, structural engineering, and intelligent infrastructure systems. The merging of ideas, new computational and manufacturing technologies, and research methods across disparate disciplines is expected to lead to the development of equitable, inclusive, and sustainable innovation and commercialization of socio-resilient infrastructure.

The future of oceans

In addition, the Signal Kinetics Group at the MIT Media Lab has joined forces with the Woods Hole Oceanographic Institution (WHOI) to spearhead Smart Oceans 2020, a series of cross-cutting, multidisciplinary virtual plenaries and workshops to be held the week of Oct. 5-9. Under the leadership of Media Lab associate professor and Doherty Chair Fadel Adib and WHOI’s assistant scientist Seth Zippel, this conference will “blue sky” the future of the ocean, aiming to accelerate the field of ocean science.

The conference will feature a mix of invited plenary speakers, lightning talks, and brainstorming sessions, all with the purpose of accelerating use-inspired convergence research. The goal of the conference organizers is to foster a wealth of synergy, connections, and cooperation, which will lead to partnerships between academia, industry, non-governmental organizations, government, philanthropy, and venture capital across sectors such as climate and environmental sustainability, computing, marine biology and ecology, geopolitics, and defense.

“This workshop is a tremendously exciting opportunity to bring together communities in the ocean sciences who don’t necessarily cross paths,” comments Lara Campbell, program director for the NSF Convergence Accelerator Program in the Office of Integrative Activities at the U.S. National Science Foundation. “More valuable still, it’s a chance to think together about what the greatest opportunities are for accelerating research into near-term impacts that allow us to sustainably engage with the tremendous resources of the oceans.”

The organizers of Smart Oceans 2020 are hopeful that by identifying convergent approaches for ocean innovation, exploration, and utilization, this series of workshops will have a direct impact on issues of sustainability, national security, and national industries. “The series of workshops will cover a variety of exciting topics, ranging from designing and building an ocean internet-of-things to identifying convergent approaches to address climate challenges,” notes Adib. “The NSF convergence accelerator is uniquely positioned to propel this field forward, as it has done for other topics through investments of over $50 million.” 

de MIT News https://ift.tt/32TdrcS

MIT undergraduates pursue research opportunities through the pandemic

Even in ordinary times, scientific process is stressful, with its demand for open-ended exploration and persistence in the face of failure. But the pandemic has added to the strain. In this new world of physical isolation, there are fewer opportunities for spontaneity and connection, and fewer distractions and events to mark the passage of time. Days pass in a numbing blur of sameness.

Working from home this summer, students participating in MIT’s Undergraduate Research Opportunities Program (UROP) did their best to overcome these challenges. Checking in with their advisors over Zoom and Slack, from as far west as Los Angeles, California and as far east as Skopje, North Macedonia, they completed two dozen projects sponsored by the MIT Quest for Intelligence. Four student projects are highlighted here.

Defending code-processing AI models against adversarial attacks 

Computer vision models have famously been fooled into classifying turtles as rifles, and planes as pigs, simply by making subtle changes to the objects and images the models are asked to interpret. But models that analyze computer code, which are a part of recent efforts to build automated tools to design programs efficiently, are also susceptible to so-called adversarial examples. 

The lab of Una-May O’Reilly, a principal research scientist at MIT, is focused on finding and fixing the weaknesses in code-processing models that can cause them to misbehave. As automated programming methods become more common, researchers are looking for ways to make this class of deep learning model more secure.

“Even small changes like giving a different name to a variable in a computer program can completely change how the model interprets the program,” says Tamara Mitrovska, a third-year student who worked on a UROP project this summer with Shashank Srikant, a graduate student in O’Reilly’s lab.

The lab is investigating two types of models used to summarize bits of a program as part of a broader effort to use machine learning to write new programs. One such model is Google’s seq2seq, originally developed for machine translation. A second is code2seq, which creates abstract representations of programs. Both are vulnerable to attacks due to a simple programming quirk: captions that let humans know what the code is doing, like assigning names to variables, give attackers an opening to exploit the model. By simply changing a variable name in a program or adding a print statement, the program may function normally, yet force the model processing it to give an incorrect answer.

This summer, from her home near Skopje, in North Macedonia, Mitrovska learned how to sift through a database of more than 100,000 programs in Java and Python and modify them algorithmically to try to fool seq2seq and code2seq. “These systems are challenging to implement,” she says. “Finding even the smallest bug can take a significant amount of time. But overall, I've been having fun and the project has been a very good learning experience for me.”

One exploit that she uncovered: Both models could be tricked by inserting “print” commands in the programs they process. That exploit, and others discovered by the lab, will be used to update the models to make them more robust.

What everyday adjectives can tell us about human reasoning

Embedded in the simplest of words are assumptions about the world that vary even among closely related languages. Take the word “biggest.” Like other superlatives in English, this adjective has no equivalent in French or Spanish. Speakers simply use the comparative form, “bigger” — plus grand in French or más grande in Spanish — to differentiate among objects of various sizes.

To understand what these words mean and how they are actually used, Helena Aparicio, formerly a postdoc at MIT and now a professor at Cornell University, devised a set of psychology experiments with MIT Associate Professor Roger Levy and Boston University Professor Elizabeth Coppock. Curtis Chen, a second-year student at MIT interested in the four topics that converge in Levy’s lab — computer science, psychology, linguistics, and cognitive science — joined on as a UROP student.

From his home in Hillsborough, New Jersey, Chen orchestrated experiments to identify why English speakers prefer superlatives in some cases and comparatives in others. He found that in scenes with more similarly sized objects, the more likely his human subjects were to prefer the word “biggest” to describe the largest object in the set. When objects appeared to fall within two clearly defined groups, subjects preferred the less-precise “bigger.” Chen also built an AI model to simulate the inferences made by his human subjects and found that it showed a similar preference for the superlative in ambiguous situations.

Designing a successful experiment can take several tries. To ensure consistency among the shapes that subjects were asked to describe, Chen generated them on the computer using HTML Canvas and JavaScript. “This way, the size differentials were exact, and we could simply report the formula used to make them,” he says.

After discovering that some subjects seemed confused by rectangle and line shapes, he replaced them with circles. He also removed the default option on his reporting scale after realizing that some subjects were using it to breeze through the tasks. Finally, he switched to the crowdsourcing platform Prolific after a number of participants on Amazon’s Mechanical Turk failed at tasks designed to ensure they were taking the experiments seriously.

“It was discouraging, but Curtis went through the process of exploring the data and figuring out what was going wrong,” says his mentor, Aparicio. 

In the end, he wound up with strong results and promising ideas for follow-up experiments this fall. “There’s still a lot to be done,” he says. “I had a lot of fun cooking up and tweaking the model, designing the experiment, and learning about this deceptively simple puzzle.”

Levy says he looks forward to the results. “Ultimately, this line of inquiry helps us understand how different vocabularies and grammatical resources of English and thousands of other languages support flexible communication by their native speakers,” he says.

Reconstructing real-world scenes from sensor data

AI systems that have become expert at sizing up scenes in photos and video may soon be able to do the same for real-world scenes. It’s a process that involves stitching together snapshots of a scene from varying viewpoints into a coherent picture. The brain performs these calculations effortlessly as we move through the world, but computers require sophisticated algorithms and extensive training. 

MIT Associate Professor Justin Solomon focuses on developing methods to help computers understand 3D environments. He and his lab look for new ways to take point cloud data gathered by sensors — essentially, reflections of infrared light bounced off the surfaces of objects — to create a holistic representation of a real-world scene. Three-dimensional scene analysis has many applications in computer graphics, but the one that drove second-year student Kevin Shao to join Solomon’s lab was its potential as a navigation tool for self-driving cars.

“Working on autonomous cars has been a childhood dream for me,” says Shao.

In the first phase of his UROP project, Shao downloaded the most important papers on 3D scene reconstruction and tried to reproduce their results. This improved his knowledge of PyTorch, the Python library that provides tools for training, testing, and evaluating models. It also gave him a deep understanding of the literature. In the second phase of the project, Shao worked with his mentor, PhD student Yue Wang, to improve on existing methods.

“Kevin implemented most of the ideas, and explained in detail why they would or wouldn’t work,” says Wang. “He didn’t give up on an idea until we had a comprehensive analysis of the problem.”

One idea they explored was the use of computer-drawn scenes to train a multi-view registration model. So far, the method works in simulation, but not on real-world scenes. Shao is now trying to incorporate real-world data to bridge the gap, and will continue the work this fall.

Wang is excited to see the results. “It sometimes takes PhD students a year to have a reasonable result,” he says. “Although we are still in the exploration phase, I think Kevin has made a successful transition from a smart student to a well-qualified researcher.”

When do infants become attuned to speech and music?

The ability to perceive speech and music has been traced to specialized parts of the brain, with infants as young as four months old showing sensitivity to speech-like sounds. MIT Professor Nancy Kanwisher and her lab are investigating how this special ear for speech and music arises in the infant brain.

Somaia Saba, a second-year student at MIT, was introduced to Kanwisher’s research last year in an intro to neuroscience class and immediately wanted to learn more. “The more I read up about cortical development, the more I realized how little we know about the development of the visual and auditory pathways,” she says. “I became very excited and met with [PhD student] Heather Kosakowski, who explained the details of her projects.”

Signing on for a project, Saba plunged into the “deep end” of cortical development research. Initially overwhelmed, she says she gained confidence through regular Zoom meetings with Kosakowski, who helped her to navigate MATLAB and other software for analyzing brain-imaging data. “Heather really helped motivate me to learn these programs quickly, which has also primed me to learn more easily in the future,” she says.

Before the pandemic shut down campus, Kanwisher’s lab collected functional magnetic resonance imaging (fMRI) data from two- to eight-week-old sleeping infants exposed to different sounds. This summer, from her home on Long Island, New York, Saba helped to analyze the data. She is now learning how to process fMRI data for awake infants, looking toward the study’s next phase. “This is a crucial and very challenging task that’s harder than processing child and adult fMRI data,” says Kosakowski. “Discovering how these specialized regions emerge in infants may be the key to unlocking mysteries about the origin of the mind.”

MIT Quest for Intelligence summer UROP projects were funded, in part, by the MIT-IBM Watson AI Lab and by Eric Schmidt, technical advisor to Alphabet Inc., and his wife, Wendy.

de MIT News https://ift.tt/2RQsWvA

The wobbling shadow of the M87* black hole

In 2019, the Event Horizon Telescope (EHT) Collaboration, including a team of MIT Haystack Observatory scientists, delivered the first image of a black hole, revealing M87* — the supermassive object in the center of the M87 galaxy. The EHT team has used the lessons learned last year to analyze the archival data sets from 2009 to 2013, some of which were not published before. The analysis reveals the behavior of the black hole image across multiple years, indicating persistence of the crescent-like shadow feature, but also variation of its orientation — the crescent appears to be wobbling. The full results appear today in The Astrophysical Journal in an article titled, “Monitoring the Morphology of M87* in 2009–2017 with the Event Horizon Telescope.”

The EHT is a global array of telescopes, performing synchronized observations using the technique of very long baseline interferometry. Together they form a virtual Earth-sized radio dish, providing a uniquely high image resolution. In 2009–13, M87* was observed by early-EHT prototype arrays, with telescopes located at three geographical sites from 2009 to 2012 and four sites in 2013. In 2017, the EHT reached maturity with telescopes located at five distinct geographical sites across the globe.

Datasets for this research were fully correlated at MIT Haystack Observatory. The 2009–2013 observations consist of less data than the ones performed in 2017, making it impossible to create an image. But the EHT team was able to use statistical modeling to look at changes in the appearance of M87* over time. In the modeling approach, the data are compared to a family of geometric templates, in this case rings of non-uniform brightness. A statistical framework is then employed to determine if the data are consistent with such models and to find the best-fitting model parameters.

“This is a beautiful example of creative data analysis. Extracting important new astrophysical understanding and squeezing new insight out of previous observations is an imaginative example of how scientists can maximally use the information content of such painstakingly collected data,” says Colin Lonsdale, director of MIT Haystack Observatory and chair of the EHT Collaboration Board. “The behavior of this event horizon scale structure over a period of years allows important additional constraints to be placed on the properties of this fascinating object.”

Expanding the analysis to the 2009–2017 observations, EHT scientists have shown that M87* adheres to theoretical expectations. The black hole’s shadow diameter has remained consistent with the prediction of Einstein’s theory of general relativity for a black hole of 6.5 billion solar masses.

“In this study, we show that the general morphology, or presence of an asymmetric ring, most likely persists on timescales of several years,” says Kazu Akiyama, research scientist at MIT Haystack Observatory and a participant in the project. “The consistency throughout multiple observational epochs gives us more confidence than ever about the nature of M87* and the origin of the shadow.”

Although the crescent diameter remained consistent, the EHT team found that the data were hiding a surprise: The ring is wobbling, and that means big news for scientists. For the first time, they can get a glimpse of the dynamical structure of the accretion flow so close to the black hole’s event horizon, in extreme gravity conditions. Studying this region holds the key to understanding phenomena such as relativistic jet launching, and will allow scientists to formulate new tests of the theory of general relativity.

The gas falling onto a black hole heats up to billions of degrees, ionizes, and becomes turbulent in the presence of magnetic fields. “Because the flow of matter is turbulent, the crescent appears to wobble with time,” says Maciek Wielgus of the Harvard and Smithsonian Center for Astrophysics, who is a Black Hole Initiative fellow, and lead author of the paper. “Actually, we see quite a lot of variation there, and not all theoretical models of accretion allow for so much wobbling. What it means is that we can start ruling out some of the models based on the observed source dynamics.”

“MIT Haystack Observatory was instrumental in organizing these early observations, correlating the massive amounts of data returned on large numbers of hard drives, and reducing the data,” says Vincent Fish, research scientist at Haystack Observatory. “While we were able to place important constraints on the size and nature of the emission in M87* at the time, the images made from the much better 2017 array data provided critical context for fully understanding what the earlier data were trying to tell us.”

Haystack scientist Geoff Crew adds, “After working on EHT technology for a decade, I’m gratified that M87* has been making equally good use of its time.”

de MIT News https://ift.tt/3mOctXl