Rachel Gordon | MIT News | CSAIL

“Foundry” tool from the Computer Science and Artificial Intelligence Lab lets you design a wide range of multi-material 3-D-printed objects.

To demonstrate Foundry, MIT researchers designed and fabricated skis with retro-reflective surfaces, a ping-pong paddle, a helmet, and even a bone that may someday be used for surgical planning. Image: Kiril Vidimče/MIT CSAIL

3-D printing has progressed over the last decade to include multi-material fabrication, enabling production of powerful, functional objects. While many advances have been made, it still has been difficult for non-programmers to create objects made of many materials (or mixtures of materials) without a more user-friendly interface.

But this week, a team from MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) will present “Foundry,” a system for custom-designing a variety of 3-D printed objects with multiple materials.

“In traditional manufacturing, objects made of different materials are manufactured via separate processes and then assembled with an adhesive or another binding process,” says PhD student Kiril Vidimče, who is first author on the paper. “Even existing multi-material 3-D printers have a similar workflow: parts are designed in traditional CAD [computer-aided-design] systems one at a time and then the print software allows the user to assign a single material to each part.”

In contrast, Foundry allows users to vary the material properties at a very fine resolution that hasn’t been possible before.

“It’s like Photoshop for 3-D materials, allowing you to design objects made of new composite materials that have the optimal mechanical, thermal, and conductive properties that you need for a given task,” says Vidimče. “You are only constrained by your creativity and your ideas on how to combine materials in novel ways.”

To demonstrate, the team designed and fabricated a ping-pong paddle, skis with retro-reflective surfaces, a tricycle wheel, a helmet, and even a bone that could someday be used for surgical planning.

Read this article on MIT News or the CSAIL website .

Redesigning multi-material objects in existing design tools would take experienced engineers many days — and some designs would actually be completely infeasible. With Foundry, you can create these designs in minutes. “

3-D printing is about more than just clicking a button and seeing the product,” Vidimče says. “It’s about printing things that can’t currently be made with traditional manufacturing.”

The paper’s co-authors include MIT Professor Wojciech Matusik and students from his Computational Fabrication Group: PhD student Alexandre Kaspar and former graduate student Ye Wang. The paper will be presented later this week at the Association for Computing Machinery’s User Interface Software and Technology Symposium (UIST) in Tokyo.

How it works

Today’s multi-material 3-D printers are mostly used for prototyping, because the materials currently used are not very functional. Users typically create preliminary models, make rapid adjustments, and then print them again. New platforms such as MIT’s MultiFab are developing highly functional materials appropriate for volume manufacturing.

Foundry, meanwhile, serves as the interface to help create such objects. To use it, you first design your object in a traditional CAD package like SolidWorks. Once the file is exported, you can determine the object’s composition by creating an “operator graph” that can include any of approximately 100 fine-tuned actions called “operators.”

Operators can “subdivide,” “remap,” or “assign” materials. Some operators cleanly divide an object into two or more different materials, while others provide more of a gradual shift from one material to another.

Foundry lets you mix and match any combination of materials and also assign specific properties to different parts of the object, combining operators together to make new ones.

For example, if you want to make a cube that is both rigid and elastic, you would assign a “rigid operator” to make one part rigid and an “elastomer operator” to the other part elastic; a third “gradient operator” connects the two and introduces a gradual transition between materials.

Users can preview their design in real-time, rather than having to wait until the final steps in the printing process to see what it will look like.

Testing the system

To test Foundry, the team tried the system on non-designers. They were given three different objects to reproduce: a teddy bear, a bone structure, and an integrated “tweel” (tire and wheel). With just an hour's explanation, users could design the bone, tire wheel, and teddy bear in an average of 56, 48, and 26 minutes, respectively.

In addition to the user study, the team also fabricated a custom wheel for a toddler tricycle. The wheel had an improved structure to maximize lateral strength, and a foam outer wheel for improved suspension.

Using Foundry to exploit the full capabilities of the 3-D printing platform enables many practical applications in medicine and more. Surgeons could create high-quality replicas of objects like bones to practice on, while doctors could also develop more comfortable dentures and other products that would benefit from having both soft and rigid components.

Vidimče’s ultimate dream is for Foundry to create a community of designers who can share new operators with each other to expand the possibilities of what can be produced. He also hopes to integrate Foundry into the workflow of existing CAD systems.

“The user should be able to iterate on the material composition in a similar manner to how they iterate on the geometry of the part being designed,” Vidimče says. “Integrating physics simulations to predict the behavior of the part will allow rapid iteration on the final design.”

The research was supported by the National Science Foundation.

Read this article on MIT News or the CSAIL website.

The 2016–2017 Undergraduate Student Advisory Group in EECS (USAGE) includes 36 students from within Course 6 who help the department grow by providing feedback about their experiences. USAGE members meet regularly with EECS Department Head Anantha Chandrakasan, Undergraduate Officer Dr. Christopher Terman, and Undergraduate Administrator Anne Hunter. Additionally, they meet with other members of the department leadership, including associate department heads Prof. David Perreault and Prof. Nancy Lynch. They will be surveying and reporting on the state of the department as it affects undergraduate students, and on how to enhance the undergraduate student community in the department. They will also discuss revitalizing the electrical engineering curriculum to encourage student interest in EE, how to provide effective online material for residential students, and other new and emerging opportunities for EECS students.

Kai Aichholz Senior, 6-2
M. Efe Akengin Junior, 6-2
Suma Anand Senior, 6-1
Anish Athalye Senior, 6-3
Roberto Brenes Senior, 6-1
Alyssa P Cartwright Senior, 6-1
William Caruso Junior, 6-3
Uttara Chakraborty Senior, 6-1
Shivani Chauhan Junior
Logan Engstrom Sophomore
Ignacio Estay Forno Junior, 6-1
Sarah Hensley Senior, 6-1
Nancy Hung Sophomore, 6-7 & 6-2
Matthew H Kalinowski Senior, 6-3
Mohamed Hassan Kané Senior, 6-3
Keertan Kini M.Eng
Isaac Kontomah Sophomore, 6-2
Natalie Lao MEng, 6-A
Allison Lemus Junior, 6-1
Jimmy Mawdsley Junior, 6-1
Remi Mir Senior, 6-3
Nathan Monroe M.Eng, 6-1
Nhat V Nguyen Sophomore, 6-3
Carla Pinzon Junior, 6-1
Daniel D Richman Senior, 6-2
Alisha Saxena Junior, 6-2
Nalini Singh Senior, 6-2
Alex Slobada Senior, 6-1
Alex Sludds Junior, 6-2
Tejas Sundaresan Senior, 6-3
Lisette Tellez Junior, 6-3
Sravya Vishnubhatla Senior, 6-3
Sarah Wooders Sophomore, 6-3
Andrew Xia Senior, 6-2
Tom Yan Senior

Meg Murphy | School of Engineering

Senior Sarah Hensley aims to improve Earth's most advanced humanoid robot, in advance of its trip to the Red Planet.

“One thing you realize working in this lab is that we are really far away from the robot apocalypse,” says MIT senior Sarah Hensley, who is working with a team of researchers to prepare Valkyrie, a humanoid robot, for future space missions. “Sometimes robots work, and sometimes they don’t. That’s our challenge.” Photo: Gretchen Ertl

Sarah Hensley is preparing an astronaut named Valkyrie for a mission to Mars. It is 6 feet tall, weighs 300 pounds, and is equipped with an extended chest cavity that makes it look distinctly female. Hensley spends much of her time this semester analyzing the movements of one of Valkyrie's arms.

As a fourth-year electrical engineering student at MIT, Hensley is working with a team of researchers to prepare Valkyrie, a humanoid robot also known as R5, for future space missions. As a teenager in New Jersey, Hensley loved to read in her downtime, particularly Isaac Asimov’s classic robot series. “I’m a huge science fiction nerd — and now I’m actually getting to work with a robot that’s real and not just in books. That’s like, wow.”

Hensley is studying Valkyrie for an advanced independent research program, or SuperUROP, as one of only three undergraduate students in the Robot Locomotion Group in MIT's Computer Science and Artificial Intelligence Laboratory. Most of her colleagues are graduate-level researchers and postdocs with extensive experience working on complex humanoids. The group is led by professor of electrical engineering and computer science Russ Tedrake, who successfully programmed Valkyrie’s predecessor (named Atlas) to open doors, turn valves, drill holes, climb stairs, and drive a car for the DARPA Robotics Challenge in 2015.

Valkyrie has 28 torque-controlled joints, four body cameras, and more than 200 individual sensors, Hensley says. The robot can walk, bend its joints, and turn a door handle. “This is one of the most advanced robots in the world. And it’s 20 feet from my desk,” she adds.

That’s largely because Valkyrie has a long way to go before it leaves for Mars. MIT is one of three institutions, including Northeastern University and the University of Edinburgh, that NASA selected to develop software enabling the robot to perform space-related tasks — open airlock hatches, attach and remove power cables, repair equipment, and retrieve samples. Oh yeah, and get to its feet when it falls down.

Hensley, who started in the lab over the summer, is intrigued by the challenge of harmonizing the movements of such a highly complex system. “I am trying to solve a very tricky problem,” she says. She’s working out how best to control Valkyrie’s elbow movements by comparing two potential approaches. One uses a main controller to gather information from the various motor systems within the arm, and then uses that data to make accurate movement decisions. The other approach is decentralized, and leaves it to each motor system to decide and act on its own.

Hensley gets animated discussing the alternatives. “Is it better to have multiple decision makers with access to different information? Or is better to have one decision maker choosing all of the motor inputs?” she asks. Hensley has already been accepted into a master’s degree program in electrical engineering at MIT. She hopes to be able to continue her work on Valkyrie.

Every day, Hensley leaves Tau Epsilon Phi, her co-ed fraternity house in the Back Bay; walks across the Massachusetts Ave. bridge to the Stata Center; and plants herself in front of two large monitors in the robotics lab. She analyzes a wealth of scientific literature and writes code for computer simulations of the equations that move the robot’s arm. Sometimes she gets up for peppermint tea, or to peer around the corner of her cubicle at Valkyrie.

One thing is certain, says Hensley. Pop culture fears that machines may soon prove superior to humans are laughable. When Valkyrie is turned on and moves, Hensley says, it often “kind of shivers and falls down. One thing you realize working in this lab is that we are really far away from the robot apocalypse,” she quips. “Sometimes robots work, and sometimes they don’t. That’s our challenge.”

Read this article on MIT News.

Sarah Goodman | MindHandHeart Initiative

Tuka Al-Hanai and Mohammad Ghassemi have connected more than 500 MIT students for more than 1,000 face-to-face lunch meetings.

MIT grad students Mohammad Ghassemi (left) and Tuka Al-Hanai started MIT Connect, a platform that matches participants for face-to-face lunches according to their background, availability, and interests. Photo: Barbara Lipohar-Staples

It started with a simple Google doc. In the spring of 2015, Mohammad Ghassemi and Tuka Al-Hanai, two graduate students in the MIT Department of Electrical Engineering and Computer Science (EECS), emailed their graduate community a sign-up sheet that read: “MIT Connect pairs members of the graduate student body for platonic, one-on-one lunches once a week over the course of the semester. This form contains a few questions to help match you with others that share your interests and schedule.”

One year, and more than 1,000 lunches later, the two students have launched the “beta edition” of MIT Connect at connected.mit.edu. Grants from the Office of the Dean for Graduate Education (ODGE), the MindHandHeart Initiative, the DeFlorez Fund, and the Legatum Fellowship have all helped Ghassemi and AlHanai to expand the program: MIT Connect is now open to graduate students, undergraduates, and postdocs, as well as all MIT alumni and employees wanting to meet others around campus. To sign up, visit the website and enter your name and MIT ID. Users can then provide their schedule and what they like to eat, and the platform tells them who to meet and when.

“We have been really surprised by the level of interest that Connect has created on campus,” Al-Hanai noted in an evaluation of the program. “We have over 500 students signed up; last fall it was featured in the campus paper; and we were even approached by an investor interested in expanding the service to the public.”

Both student founders specialize in artificial Intelligence, Ghassemi in the context of health care, and Al-Hanai in the context of speech. Their project uses an artificial intelligence algorithm they call the Maven. “Thanks to incredible feedback from users, the algorithm does a pretty good job at matching people,” says Ghassemi. Meanwhile, Al-Hanai reports that “93 percent of participants surveyed rate the program four or above, on a five-point scale, and that 52 percent made a lasting friend using the program.”

Interestingly, Connect was inadvertently inspired by a tragedy. In 2015, Ghassemi was grappling with two deaths that occurred back-to-back, in his hometown and at school. After confiding in his colleague, Al-Hanai, they started asking what personal challenges friends in their community had encountered. Ghassemi and Al-Hanai’s informal polling yielded an unexpected finding: “Time and again,” Ghassemi says, “students would explain how they wanted to find friends, mentors, or co-founders, but found mixers to be impersonal, and sometimes even awkward. One graduate student put it bluntly: ‘Nothing feels worse than going to an event alone and watching everyone else have fun.’”

After discussing with friends ways to reduce isolation and expand social networks on campus, an idea for a solution began to emerge. “One suggestion that came up was to meet with semi-random people during lunch,” says Al-Hanai. “If you were going to take time to eat lunch anyway, you might as well have lunch with someone new. An incredibly interesting person exists out there who might turn out to be your new favorite person.”

The concept moved quickly from idea to action to platform. “We quickly put together a very simple proof-of-concept and received a Graduate Student Life Grant (GSLG) from the Office of the Dean for Graduate Education (ODGE). Then, the pieces starting coming together — from a barebones Google doc to a full-fledged platform.”

For Al-Hanai, creating MIT Connect has been rewarding on many levels. “It’s wonderful to build a system from the ground up that serves the community. I’ve enjoyed learning new technologies, generating visualizations, raising support, and improving MIT Connect’s communications efforts. It’s broadened my sphere of interest. It’s a platform that I myself would use, and it’s fulfilling to hear that my peers are gaining meaningful experiences, connections, and friends through it.”

While MIT Connect has been focused on peer-to-peer pairing, the creators are currently working to launch a version of the platform to help students find mentors, employers, and entrepreneurial co-founders.

Some of the grant funding that supported the growth of MIT Connect is currently available to students, staff, and faculty at MIT: The MindHandHeart Initiative’s Innovation Fund is open for online applications through Oct. 31, and will open again in the spring.

Read this article on MIT News.

Larry Hardesty | MIT News

Algorithm could help analyze fetal scans to determine whether interventions are warranted.

A team led by Polina Golland’s group at MIT’s Computer Science and Artificial Intelligence Laboratory has made a new algorithm for identifying organs in fetal MRI scans, which should make MRI monitoring much more useful. Pictured is a stock image of a fetal MRI.

Researchers from MIT, Boston Children's Hospital, and Massachusetts General Hospital have joined forces in an ambitious new project to use magnetic resonance imaging (MRI) to evaluate the health of fetuses.

Typically, fetal development is monitored with ultrasound imaging, which is cheap and portable and can gauge blood flow through the placenta, the organ in the uterus that delivers nutrients to the fetus. But MRI could potentially measure the concentration of different chemicals in the placenta and in fetal organs, which may have more diagnostic value.

Earlier this year, in a project led by Ellen Grant’s group in the Fetal-Neonatal Neuroimaging and Developmental Science Center at Boston Children’s Hospital (BCH), members of the research team presented a paper showing that MRI measurements of oxygen absorption rates in the placenta can indicate placental disorders that might endanger the fetus. Grant is a professor of pediatrics and radiology at the Harvard Medical School (HMS).

And at the International Conference on Medical Image Computing and Computer Assisted Intervention this week, a team led by Polina Golland’s group at MIT’s Computer Science and Artificial Intelligence Laboratory presented a paper demonstrating a new algorithm for tracking organs through sequences of MRI scans, which will make MRI monitoring much more useful.

Much of Golland’s prior work has dealt with algorithmic analysis of MRI scans of the brain. “The question is, why can’t you just use everything that we’ve done in the last 25 years in the brain to apply in this case?” says Golland, a professor of electrical engineering and computer science. “And the answer is that for the brain, when the person is performing a particular task in the scanner, they’re lying still. And then after the fact, you can use algorithmic approaches to correct for very small motions. Inside the uterus, well, you can’t tell the mother not to breathe. And you can’t tell the baby not to kick.”

Frame by frame

“When you’re trying to understand whether it’s a healthy intrauterine environment, you look at fetal growth by doing measurements with ultrasound, and you look at the velocities of waveforms in the umbilical arteries,” says Grant. “But neither of those are direct measures of placental function. They’re downstream effects. Our goal is to come up with methods for assessing the spatiotemporal function of the placenta directly. If you really want to intervene, you want to intervene before the placenta fails.”

Grant leads the clinical arm of the project together with Lawrence Wald, a physicist at Massachusetts General Hospital and a professor of radiology at HMS. Elfar Adalsteinsson, an MIT professor of electrical engineering and computer science, is collaborating with colleagues at BCH to develop new MRI technologies for fetal imaging, and Golland’s group is in charge of developing software for interpreting the images.

An MRI image might consist of hundreds of two-dimensional cross sections of an anatomical region, stitched into a three-dimensional whole. Measuring chemical changes over time requires analyzing sequences of such three-dimensional representations — about 300, in the case of the new paper. The researchers refer to each MRI image in a series as a “frame,” analogous to frames of video.

The first step in localizing chemical changes to particular organs, of course, is identifying the organs. That’s where the researchers’ new algorithm comes in.

With MRI images of brain activity, it’s comparatively easy to determine which anatomical features in one frame correspond to which features in the next. The subject’s head is immobilized, and brain regions don’t change shape or location over the course of a scan. Algorithmically, the standard method for coordinating frames is to identify a region in the first frame and then map it separately onto each of the frames that follow.

With fetal MRIs, that won’t work, because the fetus may have moved dramatically between, say, frame one and frame 200. So Golland and her co-authors — including first author Ruizhi Liao, an MIT graduate student in electrical engineering and computer science; Grant; and Adalsteinsson — took a different approach.

On a roll

Their algorithm begins by finding a mathematical function that maps the pixels of the first frame onto those of the second; then it maps the mathematically transformed version of the first frame onto the third, and so on. The end result is a series of mathematical operations that describes the evolution of the scan as a whole. “The way to think about how this algorithm works is, it takes the baseline frame — for example, the first one — and it rolls it down the sequence,” Golland says.

Next, a human expert draws very precise boundaries around the elements of interest in the first frame — in this case, not just the placenta but the brain and liver as well. Those elements’ movements or deformations from frame to frame can then be calculated using the previously determined mathematical operations.

Hand-drawing organ boundaries — or “segmenting” an MRI scan — is a time-consuming process. But performing it only once is much less onerous than performing it 300 times.

In order to evaluate the accuracy of their algorithm, the researchers hand-segmented an additional five frames. “Two members of the team sat there for about a week and drew outlines,” Golland says. “It’s a very painful validation process, but you have to do it to believe the results.” The algorithm’s segmentations accorded very well with those performed by hand.

“One of the big problems in high-speed acquisition and MR [magnetic resonance] acquisition is definitely the incorporation of motion and trying to deal with motion issues,” says Sarang Joshi, a professor of bioengineering at the University of Utah. “Modeling and incorporating the deformation estimation in MR acquisition is a big challenge, and we have been working on it as well, and many other people have been working on it. So this is a really great step forward.”

Read this article on MIT News.

Larry Hardesty | MIT News

With new algorithms, data scientists could accomplish in days what has traditionally taken months.

“The goal of all this is to present the interesting stuff to the data scientists so that they can more quickly address all these new data sets that are coming in,” says Max Kanter MEng ’15.

Last year, MIT researchers presented a system that automated a crucial step in big-data analysis: the selection of a “feature set,” or aspects of the data that are useful for making predictions. The researchers entered the system in several data science contests, where it outperformed most of the human competitors and took only hours instead of months to perform its analyses.

This week, in a pair of papers at the IEEE International Conference on Data Science and Advanced Analytics, the team described an approach to automating most of the rest of the process of big-data analysis — the preparation of the data for analysis and even the specification of problems that the analysis might be able to solve.

The researchers believe that, again, their systems could perform in days tasks that used to take data scientists months.

“The goal of all this is to present the interesting stuff to the data scientists so that they can more quickly address all these new data sets that are coming in,” says Max Kanter MEng ’15, who is first author on last year’s paper and one of this year’s papers. “[Data scientists want to know], ‘Why don’t you show me the top 10 things that I can do the best, and then I’ll dig down into those?’ So [these methods are] shrinking the time between getting a data set and actually producing value out of it.”

Both papers focus on time-varying data, which reflects observations made over time, and they assume that the goal of analysis is to produce a probabilistic model that will predict future events on the basis of current observations.

Real-world problems

The first paper describes a general framework for analyzing time-varying data. It splits the analytic process into three stages: labeling the data, or categorizing salient data points so they can be fed to a machine-learning system; segmenting the data, or determining which time sequences of data points are relevant to which problems; and “featurizing” the data, the step performed by the system the researchers presented last year.

The second paper describes a new language for describing data-analysis problems and a set of algorithms that automatically recombine data in different ways, to determine what types of prediction problems the data might be useful for solving.

According to Kalyan Veeramachaneni, a principal research scientist at MIT’s Laboratory for Information and Decision Systems and senior author on all three papers, the work grew out of his team’s experience with real data-analysis problems brought to it by industry researchers.

“Our experience was, when we got the data, the domain experts and data scientists sat around the table for a couple months to define a prediction problem,” he says. “The reason I think that people did that is they knew that the label-segment-featurize process takes six to eight months. So we better define a good prediction problem to even start that process.”

In 2015, after completing his master’s, Kanter joined Veeramachaneni’s group as a researcher. Then, in the fall of 2015, Kanter and Veeramachaneni founded a company called Feature Labs to commercialize their data-analysis technology. Kanter is now the company’s CEO, and after receiving his master’s in 2016, another master’s student in Veeramachaneni’s group, Benjamin Schreck, joined the company as chief data scientist.

Data preparation

Developed by Schreck and Veeramachaneni, the new language, dubbed Trane, should reduce the time it takes data scientists to define good prediction problems, from months to days. Kanter, Veeramachaneni, and another Feature Labs employee, Owen Gillespie, have also devised a method that should do the same for the label-segment-featurize (LSF) process.

To get a sense of what labeling and segmentation entails, suppose that a data scientist is presented with electroencephalogram (EEG) data for several patients with epilepsy and asked to identify patterns in the data that might signal the onset of seizures.

The first step is to identify the EEG spikes that indicate seizures. The next is to extract a segment of the EEG signal that precedes each seizure. For purposes of comparison, “normal” segments of the signal — segments of similar length but far removed from seizures — should also be extracted. The segments are then labeled as either preceding a seizure or not, information that a machine-learning algorithm can use to identify patterns that indicate seizure onset.

In their LSF paper, Kanter, Veeramachaneni, and Gillespie define a general mathematical framework for describing such labeling and segmentation problems. Rather than EEG readings, for instance, the data might be the purchases by customers of a particular company, and the problem might be to determine from a customer’s buying history whether he or she is likely to buy a new product.

There, the pertinent data, for predictive purposes, may be not a customer’s behavior over some time span, but information about his or her three most recent purchases, whenever they occurred. The framework is flexible enough to accommodate such different specifications. But once those specifications are made, the researchers’ algorithm performs the corresponding segmentation and labeling automatically.

Finding problems

With Trane, time-series data is represented in tables, where the columns contain measurements and the times at which they were made. Schreck and Veeramachaneni defined a small set of operations that can be performed on either columns or rows. A row operation is something like determining whether a measurement in one row is greater than some threshold number, or raising it to particular power. A column operation is something like taking the differences between successive measurements in a column, or summing all the measurements, or taking just the first or last one.

Fed a table of data, Trane exhaustively iterates through combinations of such operations, enumerating a huge number of potential questions that can be asked of the data — whether, for instance, the differences between measurements in successive rows ever exceeds a particular value, or whether there are any rows for which it is true that the square of the data equals a particular number.

To test Trane’s utility, the researchers considered a suite of questions that data scientists had posed about roughly 60 real data sets. They limited the number of sequential operations that Trane could perform on the data to five, and those operations were drawn from a set of only six row operations and 11 column operations. Remarkably, that comparatively limited set was enough to reproduce every question that researchers had in fact posed — in addition to hundreds of others that they hadn’t.

“Probably the biggest thing here is that it’s a big step toward enabling us to represent prediction problems in a standard way so that you could share that with other analysts in an abstraction from the problem specifics,” says Kiri Wagstaff, a senior researcher in artificial intelligence and machine learning at NASA’s Jet Propulsion Laboratory. “What I would hope is that this could lead to improved collaboration between whatever domain experts you’re working with and the data analysts. Because now the domain experts, if they could learn and would be willing to use this language, could specify their problems in a much more precise way than they’re currently able to do.”

Read this article on MIT News.

Rachel Gordon | CSAIL

This week a team from CSAIL’s computer vision group co-hosted the first Scene Parsing Challenge at the 2016 European Conference on Computer Vision (ECCV) in Amsterdam. 

The ground-truth images and the scene parsing results given by the top three algorithms.

This week a team from CSAIL’s computer vision group co-hosted the first Scene Parsing Challenge at the 2016 European Conference on Computer Vision (ECCV) in Amsterdam. The challenge was focused on scene recognition, and using data to enable algorithms to classify and segment objects in scenes.

Scene recognition has important applications in robotics and even psychology. Better algorithms could determine actions happening in a given environment, spot inconsistent objects or human behaviors, and even predict future events.

“This challenge creates a standard benchmark for model comparison and motivates researchers to propose stronger models with improved performances in scene parsing and classification.” said CSAIL PhD Bolei Zhou, who oversaw both of the scene understanding challenges.

The event was jointly held with the Large Scale Visual Recognition Challenge 2016. Among the other researchers who helped organize the challenge include PhD students Hang Zhao and Xavier Puig, MIT professor Antonio Torralba and professor Sanja Fidler from the University of Toronto.

Tackling the challenge

For the the scene parsing challenge, groups were given fully annotated images to train their algorithms on.

They were then tested on how well their algorithms segmented and classified pixels (of new photos), in terms of pixel accuracy and ratios of objects in the scene.

For example, in a photo of a building with trees and cars, the algorithm might segment the “building” image region with a 60% pixel accuracy, and 60 percent ratio of the total photo. The “car” image region might be a 30% pixel accuracy, at 30 percent, and the “tree” image region at 10% pixel accuracy with a 10 percent ratio of the total photo.

To train and test their algorithms participants used CSAIL’s ADE20K dataset of more than 22,000 photos, annotated with objects and parts. Algorithms were tested on 150 different semantic categories that include “sky”, “road”, “grass”, and more discrete object types like “person”, “car”, and “bed.”

From a total of 75 entries submitted by 22 teams from around the world, the winning team had a score of 0.57205, or 57 percent. In other words, for any given image the algorithm was accurate for more than half of the picture’s pixels.

Top ranked teams were invited to present their algorithms at ECCV’16 in Amsterdam, where more than two thousand researchers worldwide attended.

A live demo of scene parsing is here.

Places Scene Classification Challenge

The team, which included recent PhD graduate Aditya Khosla and principal research scientist Aude Oliva, simultaneously hosted the second “Places Scene Classification Challenge,” where the goal was to identify the scene category in a given photo.

For the challenge the team created the Places2 dataset, which has more than ten million images and over 400 distinct scene categories.

The algorithms produced a list of five possible categories, which were evaluated on the label that best matched the “human” assigned choice. Since many environments can be described using different words (i.e. a gym might also be “fitness center”), the idea was to allow the algorithm to identify multiple scene categories.

The winning team’s algorithms represented a marked improvement over previous algorithms for scene classification.

“In many ways, hosting a public challenge is as important as publishing seminal work,” said Oliva.

Scene classification example. A live demo is here.

In the field of deep learning, data has become the underpinning for making advances. Large labeled datasets such as Pascal, COCO and ImageNet have allowed machine learning algorithms to reach near-human semantic classification of visual patterns like objects and scenes.

“We released larger and richer data so that researchers could not only work together, but challenge each other and jointly push the slope of the field,” said Zhou. “It is like the Olympic Games in the computer vision field.”

Read this article on the CSAIL website.

Expert in optimization theory to join department leadership.

Asu Ozdaglar, the Joseph F. and Nancy P. Keithley Professor of Electrical Engineering, has been appointed as associate head of the Department of Electrical Engineering and Computer Science, effective January 1, 2017. Ozdaglar succeeds David Perreault, professor of electrical engineering, who has served in the role since November 2013.

Ozdaglar is best known for her contributions in the areas of optimization theory, economic and social networked systems, and game theory. She has made several key contributions to optimization theory, ranging from convex analysis and duality to distributed and incremental algorithms for large scale systems and data processing. She co-authored the book entitled “Convex Analysis and Optimization” with Dimitri Bertsekas and Angelia Nedich. Ozdaglar has focused a large part of her research on integrating analysis of social and economic interactions into the study of networks. Her work spans many dimensions of this area, including analysis of learning and communication, diffusion and information propagation, influence in social networks, and study of cascades and systemic risk in economic and financial systems. She continues to make key contributions to game theory, including learning dynamics and computation of Nash equilibria.

She has also organized numerous conferences and sessions on game theory, networks, and distributed optimization. Ozdaglar has been recognized by these communities as a leader, and has won several key awards, including the prestigious Donald P. Eckman Award of the American Automatic Control Council. She was also recognized at MIT with the award of the inaugural Steven and Renee Finn Faculty Research Innovation Fellowship.

Ozdaglar’s numerous educational contributions to the MIT community include the development of several graduate and undergraduate courses. She developed a graduate-level game theory subject (6.254, Game Theory with Engineering Applications), and co-developed an undergraduate course on networks (6.207, Networks), which is jointly listed with Economics. For her educational contributions, she was honored with the MIT Graduate Student Council's Teaching Award for the School of Engineering in 2004.

In October 2014, Ozdaglar became the Director of the Laboratory for Information and Decision Systems (LIDS) and the Associate Director of the Institute for Data, Systems, and Society (IDSS). She is playing a critical leadership role in shaping the future of IDSS. Ozdaglar was also a Technical Program Co-Chair of the 2015 Rising Stars program in EECS.

Perreault will continue to serve as associate department head through December 2016. Perreault spearheaded many efforts aimed at improving the undergraduate curriculum during his term in the role, including streamlining the 6/8 double major, helping with the new curriculum, organizing trips for EE students to industry (EExplore), and ongoing efforts to revitalize the electrical engineering undergraduate curriculum.

“I would like to extend my sincere thanks to Dave and express my appreciation for his ongoing service,” said EECS Department Head Anantha Chandrakasan, the Vannevar Bush Professor of Electrical Engineering and Computer Science. “I know we will continue to benefit from his extraordinary leadership and initiative as associate department head.”

Eric Brown | MIT Industrial Liaison Program

James Kirtley discusses the transition from gas to electric motors and the impact these motors have had on modern technologies.

Professor James Kirtley Photo: David Sella

Electric motors have been around since Thomas Davenport built the first functional model in 1834, and they have played a growing part in our lives ever since. Today, they continue to replace diesel and gas engines, as well as hydraulic cylinders, while evolving into new designs optimized for robots and other technologies.

“Something like 40 percent of electric power is used to drive motors, and that number will only grow,” says James Kirtley, a professor in MIT’s Department of Electrical Engineering and Computer Science and in MIT's Research Laboratory of Electronics. “Electric motors are being used more widely in ships, airplanes, trains, and cars. We’re also seeing a lot more electric motors in robots.”

The ongoing transition from gas to electric is primarily driven by the need for more efficient devices that run with cleaner energy sources. Yet, electric motors also tend to be more responsive, and are more adaptable to new applications, especially in smaller devices.

As one of the world’s leading experts on electric motors, Kirtley’s philosophy is that one size does not fit all. “If you take into account the specific application, you can build a motor that is far better adapted than a general purpose motor,” says Kirtley. “For example, I’m working with someone who is making robots for medical assist, and he needs motors with very special characteristics.”

Many of the newer types of electric motors tend to be much smaller and run on less power than in the past. “I started working with electric motors 40 years ago designing big nuclear generators with 20-foot long rotors that were 6 feet in diameter and could produce a gigawatt and a half of power,” says Kirtley. “I’m now building motors on the order of 100 to 200 Watts for appliance motors, which are kinder and gentler to the system powering them. In the automotive industry, the average automobile has dozens of small motors for things like door locks, wiper blades, air conditioning, and seat positioners.”

Even excluding the separate field of MEMS (micro-electromechanical systems), which Kirtley is not directly involved in, electric motors are now shrinking to as small as the 1 W devices found in cellphones. A variety of innovative new compact motors are being developed all over MIT, says Kirtley, who points to an interesting variable reluctance motor for a prosthetic foot being designed at MIT's Center for Art, Science and Technology.

Clean transportation is another source of innovation in electric motors. “In my lab we’re doing work with a small company in Cambridge that makes bicycle assist wheels,” says Kirtley. “The wheel stores some energy, and can react to pedaling forces to help it climb hills. These are entirely new applications.”

Improving microgrids with smart motors

Kirtley’s early involvement with power-plant generators led him to study electric power systems. His research into the subject culminated in 2010 with his Wiley-published book, “Electric Power Principles.” Lately, he’s been focusing on the customer-end of the system, where he is finding a role for electric motors in helping distribution systems adapt to intermittent, user-generated solar power.

“Electric power distribution systems are being stretched by the growing use of distributed renewable generation, such as rooftop solar,” says Kirtley. “Typically, electricity is transmitted from large power plants through extra high voltage wires, and the voltage is stepped down and delivered to customers. The problem with rooftop solar is that it looks to the power system like a reduction in load, but as solar cells become more widespread, homes will at times be able to produce more power than they’re consuming. So the power flows backward, which makes everything more complicated.”

Utilities are now working on smart microgrids that can adapt to distributed, multidirectional power. The greater flexibility is primarily enabled by smarter electronics, as well as efficient, distributed battery storage. Yet, microgrids still have a problem with even brief power outages, which can sometimes cause them to shut down.

“We’re thinking about the dynamics of motors connected to microgrids, and how you can improve their stability and make them work better,” Kirtley says. “We’re concerned about continuity of supply, which is especially important with microgrids supporting large server farms. You don’t want your system to be forced into an involuntary reboot simply because you had a glitch in the electric power supply. Electric motors can add more reliability.”

Electrifying the Cheetah robot

Many of the recent innovations in electric motors are found in robotics, which need smarter motors that can reliably deliver variable levels of power on demand for short periods. Electric motors can provide mobile robots with significantly longer battery life compared to traditional hydraulic systems. “Hydraulics are controlled primarily through throttle valves, so a lot of energy is wasted pumping and controlling the hydraulic fluid,” says Kirtley.

Kirtley has been working with Professor Jeffrey Lang on developing customized electric motors for Department of Mechanical Engineering colleague Sangbae Kim’s robotic cheetah, a running, jumping quadruped that has gained widespread publicity in recent years. The cheetah’s new motor is not only more efficient, but also more powerful, although only in short intervals.

“A secondary advantage of electric motors is responsiveness and control,” says Kirtley. “We can build a motor that can produce considerable torque in short spikes, even if we can’t necessarily produce the forces for a long period of time. It’s perhaps a little too powerful for the cheetah, which can now jump so high in the air, it probably wouldn’t survive the landing if they didn’t catch it.”

Electric motors transform ships

Most commercial ships still use diesel engines, while many naval vessels use gas turbine engines. Yet shipping is quickly moving toward electric propulsion, motivated primarily by efficiency, says Kirtley. Aside from nuclear-powered vessels, these tend to be hybrid systems in which diesel or gas generators drive an electric motor.

“A traditional gear drive for ships has some very decided disadvantages,” he says. “For example, most destroyers in the U.S. Navy have gearboxes with very precise machining requirements, and are therefore expensive. They also require a fixed gear ratio between the engine and the water, so the prime mover is not operating near its peak fuel efficiency. Because of that, much of commercial shipping, and virtually all cruise ships, are now moving to electric propulsion, and even the U.S. Navy is starting to use it for its latest destroyer.”

The other problem with fuel-driven gear-drive engines is “the tyranny of the shaft line,” says Kirtley. “If you’re going to use a direct gear drive, the engine, gearbox, and propeller must line up very precisely, which often takes up valuable real estate within the body of the ship. With electric propulsion, we don’t have that problem. The engines can be placed anywhere where it’s convenient. In cruise ships, for example, the motors are place in a pod underneath the ship.”

Electric cars: adapting the engine to the road

The main difference between electric propulsion on ships vs. cars is related to torque requirements, says Kirtley. “On a ship, top speed defines the torque requirement,” he explains. “Automobile propulsion occurs across a speed range of about 10 to 1, with an engine that idles at 600 rpm capable of redlining at about 6,000 rpm. The best motors for cars are those that are adaptable to a very wide speed range.”

The variable speeds used in a car require that “the gearbox adapt the engine to the road,” says Kirtley. “You can generate an electric motor that can propel an automobile without a gear shift.”

In the past, Kirtley has consulted with Tesla Motors on its electric cars, and both agree that “the induction motor is the best for electric automobiles,” says Kirtley. Many other electric car manufacturers are still using permanent magnet motors, which he says are intrinsically less efficient in the wide speed and torque range required by car propulsion.

“For any given electric motor there is a tradeoff between excitation — making the operating magnetic field within the machine — and reaction, providing current to push on that exciting field inside the machine,” explains Kirtley. “In a permanent magnet machine that field is constant and cannot be adjusted, so a machine that is turning very fast but making relatively little torque is dropping a lot of power in losses in the machine’s magnetic iron. In an induction motor you can back off on the excitation to provide torque at the energetically optimal fashion. You can improve the drive efficiency of a car over a complete drive cycle by as much of a factor of two in fewer losses.”

Induction motors aren’t optimal for all applications, however, which brings Kirtley back to his main thesis: “In the development of motors for modern applications, it is most important to understand the totality of operational requirements,” he says. “That is key to making electric motors that will accomplish what they do best: provide motion in a responsive and efficient fashion.”

Read this article on MIT News.

Larry Hardesty | MIT News

New model-fitting technique is efficient even for data sets with hundreds of variables.

A team, including researchers from MIT’s Computer Science and Artificial Intelligence Laboratory, has created a new set of algorithms that can efficiently fit probability distributions to high-dimensional data. Image: MIT News

Data analysis — and particularly big-data analysis — is often a matter of fitting data to some sort of mathematical model. The most familiar example of this might be linear regression, which finds a line that approximates a distribution of data points. But fitting data to probability distributions, such as the familiar bell curve, is just as common.

If, however, a data set has just a few corrupted entries — say, outlandishly improbable measurements — standard data-fitting techniques can break down. This problem becomes much more acute with high-dimensional data, or data with many variables, which is ubiquitous in the digital age.

Since the early 1960s, it’s been known that there are algorithms for weeding corruptions out of high-dimensional data, but none of the algorithms proposed in the past 50 years are practical when the variable count gets above, say, 12.

That’s about to change. Earlier this month, at the IEEE Symposium on Foundations of Computer Science, a team of researchers from MIT’s Computer Science and Artificial Intelligence Laboratory, the University of Southern California, and the University of California at San Diego presented a new set of algorithms that can efficiently fit probability distributions to high-dimensional data.

Remarkably, at the same conference, researchers from Georgia Tech presented a very similar algorithm.

The pioneering work on “robust statistics,” or statistical methods that can tolerate corrupted data, was done by statisticians, but both new papers come from groups of computer scientists. That probably reflects a shift of attention within the field, toward the computational efficiency of model-fitting techniques.

“From the vantage point of theoretical computer science, it’s much more apparent how rare it is for a problem to be efficiently solvable,” says Ankur Moitra, the Rockwell International Career Development Assistant Professor of Mathematics at MIT and one of the leaders of the MIT-USC-UCSD project. “If you start off with some hypothetical thing — ‘Man, I wish I could do this. If I could, it would be robust’ — you’re going to have a bad time, because it will be inefficient. You should start off with the things that you know that you can efficiently do, and figure out how to piece them together to get robustness.”

Resisting corruption

To understand the principle behind robust statistics, Moitra explains, consider the normal distribution — the bell curve, or in mathematical parlance, the one-dimensional Gaussian distribution. The one-dimensional Gaussian is completely described by two parameters: the mean, or average, value of the data, and the variance, which is a measure of how quickly the data spreads out around the mean.

If the data in a data set — say, people’s heights in a given population — is well-described by a Gaussian distribution, then the mean is just the arithmetic average. But suppose you have a data set consisting of height measurements of 100 women, and while most of them cluster around 64 inches — some a little higher, some a little lower — one of them, for some reason, is 1,000 inches. Taking the arithmetic average will peg a woman’s mean height at 6 feet 4 inches, not 5 feet 4 inches.

One way to avoid such a nonsensical result is to estimate the mean, not by taking the numerical average of the data, but by finding its median value. This would involve listing all the 100 measurements in order, from smallest to highest, and taking the 50th or 51st. An algorithm that uses the median to estimate the mean is thus more robust, meaning it’s less responsive to corrupted data, than one that uses the average. The median is just an approximation of the mean, however, and the accuracy of the approximation decreases rapidly with more variables. Big-data analysis might require examining thousands or even millions of variables; in such cases, approximating the mean with the median would often yield unusable results.

Identifying outliers

One way to weed corrupted data out of a high-dimensional data set is to take 2-D cross sections of the graph of the data and see whether they look like Gaussian distributions. If they don’t, you may have located a cluster of spurious data points, such as that 80-foot-tall woman, which can simply be excised.

The problem is that, with all previously known algorithms that adopted this approach, the number of cross sections required to find corrupted data was an exponential function of the number of dimensions. By contrast, Moitra and his coauthors — Gautam Kamath and Jerry Li, both MIT graduate students in electrical engineering and computer science; Ilias Diakonikolas and Alistair Stewart of USC; and Daniel Kane of USCD — found an algorithm whose running time increases with the number of data dimensions at a much more reasonable rate (or, polynomially, in computer science jargon).

Their algorithm relies on two insights. The first is what metric to use when measuring how far away a data set is from a range of distributions with approximately the same shape. That allows them to tell when they’ve winnowed out enough corrupted data to permit a good fit.

The other is how to identify the regions of data in which to begin taking cross sections. For that, the researchers rely on something called the kurtosis of a distribution, which measures the size of its tails, or the rate at which the concentration of data decreases far from the mean. Again, there are multiple ways to infer kurtosis from data samples, and selecting the right one is central to the algorithm’s efficiency.

The researchers’ approach works with Gaussian distributions, certain combinations of Gaussian distributions, another common distribution called the product distribution, and certain combinations of product distributions. Although they believe that their approach can be extended to other types of distributions, in ongoing work, their chief focus is on applying their techniques to real-world data.

Read this article on MIT News.

Adam Connor-Simons | CSAIL

CSAIL team's “Gitless” fixes Git’s biggest issues — and suggests their approach could help improve systems like Gmail and Dropbox.

“Gitless" removes complicated concepts like "staging" and "stashing," without fundamentally changing Git's core functionality. Image: Santiago Perez De Rosso

Git is an open-source system with a polarizing reputation among programmers. It’s a powerful tool to help developers track changes to code, but many view it as prohibitively difficult to use.

To make it more user-friendly, a team from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) has developed “Gitless,” an interface that fixes many of the system’s core problems without fundamentally changing what it does.

“With Gitless we’ve developed a tool that we think is easier to learn and use, but that still keeps the core elements that make Git popular,” says graduate student Santiago Perez De Rosso, who co-wrote a related paper with MIT Professor Daniel Jackson. “What’s particularly encouraging about this work is that it suggests that the same approach might be used to improve the usability of other software systems, such as Dropbox and Google Inbox.”

Gitless was developed, in part, by looking at nearly 2,400 Git-related questions from the popular programming site StackOverflow. The team then outlined some of Git’s biggest issues, including its concepts of “staging” and “stashing,” and proposed changes aimed at minimizing those problems.

Because Gitless is implemented on top of Git, users can easily switch between the two without having migrate code from one to the other. Plus, their collaborators don’t even have to know that they aren’t big fans of Git.

Perez De Rosso will present the paper at next month’s ACM SIGPLAN conference on “Systems, Programming, Languages and Applications: Software for Humanity” in Amsterdam.

How it works

Git is what’s called a “version control system.” It allows multiple programmers to track changes to code, including making “branches” of a file that can be worked on individually.

Users make changes and then save (or “commit”) them so that everyone knows who did what. If you and a colleague are on version 10 of a file, and you want to try something new, you can create a separate “branch” while your friend works on the “master.”

Makes sense, right? But things get confusing quickly. One feature of Gitless is that it eliminates “staging,” which lets you save just certain parts of a file. For example, let’s say you have a file with both finished and unfinished changes, and you’d like to commit the finished changes. “Staging” lets you commit those changes while keeping the others as a work-in-progress.

However, having a file with both a staged and working version creates tricky situations. If you stage a file and make more changes that you then commit, the version that’s committed is the one you staged before, not the one you’re working on now.

Gitless essentially hides the staging area altogether, which makes the process much clearer and less complex for the user. Instead, there’s a much more flexible “commit” command that still allows you to do things like selecting segments of code to commit.

Another concept that Gitless removes is “stashing.” Imagine that you’re in the middle of a project and have to switch to a different branch of it, but don’t yet want to commit your half-done work. Stashing takes the changes you’ve made and saves them on a stack of unfinished changes that you can restore later. (The key difference between stashing and staging is that, with stashing, changes disappear from the working directory.)

“The problem is that, when switching branches, it can be hard to remember which stash goes where,” says Perez De Rosso. “On top of that, stashing doesn’t help if you are in the middle of an action like a merge that involves conflicting files.”

Gitless solves this issue by making branches completely independent from each other. This makes it much easier and less confusing for developers who have to constantly switch between tasks.

Gitless certainly isn’t the first effort to improve Git. But according to Philip Guo, an assistant professor of cognitive science at the University of California at San Diego, who was not involved in the project, it is the first to go beyond Git’s interface and actually deal with core conceptual issues.

“This work applies rigorous software-design research techniques to uncover shortcomings in one of the world’s most widely-used pieces of software,” Guo says. “In the past, many practitioners have made anecdotal arguments both for and against Git, but no prior work has taken a scientific approach to unpacking those arguments.”

Results

The team also conducted a user study to test Gitless’ performance against Git. The researchers found that Gitless users were more successful at completing tasks than Git users, and, for at least one task, performed it significantly quicker. (Perez De Rosso points out that the study’s participants were all well-versed in Git, and suggest that the results may have been even more pronounced if the team had tested Gitless on people with no Git experience.)

In a post-task survey, participants were particularly impressed with Gitless’ ability to transition between branches, which they described as “very smooth” and “way more intuitive.” Guo describes Gitless as a valuable form of “training wheels” to help beginner programmers get started with Git. At a higher level, he says that the team’s framework could be an important tool for looking at other software systems.

Perez De Rosso says that he is particularly excited by the possibility of analyzing Google Inbox’s concept of bundled “conversations,” as well as Dropbox’s notion of “shared folders.” “Perhaps the most long-lived contribution of this research is not the analysis of Git itself, but rather the methodology that the authors used to analyze, dissect, and redesign a popular piece of software,” Guo says. “The power of the authors' approach to analyzing design flaws can be applied to many kinds of popular software.”

Read this article on MIT News.

Rob Matheson | MIT News

The Engine will provide funding, space, and expertise — powering a network of innovation networks.

Today MIT President L. Rafael Reif announced the creation of The Engine, a new kind of enterprise designed to support startup companies working on scientific and technological innovation with the potential for transformative societal impact.

President Reif made the announcement at an evening event at The Engine’s Central Square, Cambridge, Massachusetts, headquarters attended by entrepreneurs, business leaders, investors, and members of the MIT community.

The Engine is designed to meet an underserved need. In Kendall Square and Greater Boston, many breakthrough innovations cannot effectively leave the lab because companies pursuing capital- and time-intensive technologies have difficulty finding stable support and access to the resources they need.

MIT's new venture, The Engine, will provide funding, space, and expertise — powering a network of innovation networks. MIT Video Productions

“If we hope for serious solutions to the world’s great challenges, we need to make sure the innovators working on those problems see a realistic pathway to the marketplace,” President Reif says. “The Engine can provide that pathway by prioritizing breakthrough ideas over early profit, helping to shorten the time it takes these startups to become ‘VC-ready,’ providing comprehensive support in the meantime, and creating an enthusiastic community of inventors and supporters who share a focus on making a better world. We believe this approach can offer exponential growth to regions that pursue it successfully — and we want Greater Boston to lead the way.”

To fuel The Engine, MIT will seek to attract hundreds of millions of dollars of support and to make available, for entrepreneurs, hundreds of thousands of square feet of space in Kendall Square and nearby communities.

The Engine will also introduce startups to their entrepreneurial peers and to established companies, in innovation clusters across the region and around the world: It seeks to power a network of innovation networks.

“The Engine builds on work MIT has undertaken in recent years to stoke innovation on and near our campus — including starting up the MIT Innovation Initiative in 2014,” says MIT Provost Martin Schmidt. “Our faculty, alumni, and student entrepreneurs directly serve the Institute’s mission of using science and technology to make a better world, because the problems they pursue tend to be the hardest ones they can find.”

The Engine, whose launch has been led by MIT Executive Vice President and Treasurer Israel Ruiz, will offer startups “patient” capital; affordable workspaces; access to specialized equipment; streamlined business services; and a community of like-minded entrepreneurs.

“We want highly disruptive entrepreneurs to stay in Greater Boston,” Ruiz says. “This is where the boldest ideas in the world should find their home.”

A startup as different as the startups it will serve

The Engine seeks eventually to support in steady state 60 locally based startups, primarily those that are developing “tough” technologies — breakthrough ideas that require time and patient capital to commercialize — in a range of sectors including biotechnology, robotics, manufacturing, medical devices, and energy.

Accepted startups will participate in The Engine’s incubator for up to 12 months. In that time, they will receive financial investments as well as guidance in business planning and access to shared services such as legal, technology licensing, and administrative assistance. Entrepreneurs will be able to take advantage of specialized equipment, services, expertise, and space through an online marketplace developed for The Engine.

To financially support startups, The Engine will form a venture-investing arm, which will provide significant, long-term capital support to help startups transition from ideation phases to commercial success. The Engine venture funds will demand less equity in startups than is typical, allowing founders to maintain more control over their companies. The Engine is also actively exploring avenues to support nonprofit startups.

To start, The Engine is raising $150 million for a first fund, with $25 million coming from MIT as a limited partner.

The Engine will also provide space to participating startups. Its Cambridge headquarters will offer 26,000 square feet, and MIT seeks to make available in Kendall Square and nearby neighborhoods, as soon as possible, a first wave of additional space, so that participating companies could soon have over 200,000 square feet designed to meet their needs.

As part of The Engine, MIT is starting a pilot program with the City of Cambridge — “Pathways to Invention” — designed to give Cambridge schoolchildren hands-on experiences, at MIT and around the city, that introduce them to the work of invention and to the college and career paths that lead to it.

Acceleration at two vital stages for startups

The Engine will support entrepreneurs during two specific stages of innovation: the early “proof of product” stage, when entrepreneurs initially translate a novel idea into a commercializable venture, and the later stage between advanced prototypes and commercial production.

For the initial stage — which includes initial prototyping and product testing — The Engine will make meaningful investments that will allow participating startups to get off the ground quickly and operate long enough to prove their concepts’ viability. It will also provide shared access to costly infrastructure and resources and foster a community of experts and innovators who can lend expertise to the startups: Building upon lessons learned by the Innovation Initiative in supporting innovation and entrepreneurship on campus, MIT has identified many of the challenges that startups face as they enter these earliest stages. For follow-up stages — after the startups have advanced prototypes but haven’t yet demonstrated their products at commercial scale — it intends to provide increased support to promising startups.

The Engine will establish a network of existing and forthcoming shared spaces surrounding the MIT campus — including offices, labs, and prototyping and maker spaces — that will enable convenient and cost-efficient sharing of capital-intensive equipment, which is vital to startups aiming to make physical products.

Further linking the new facilities, Engine-sponsored transportation will help entrepreneurs and others move among facilities in Cambridge, Boston’s Seaport District, spaces in MIT’s West Campus, and other nearby areas featuring resources that will best serve the needs of entrepreneurs.

A marketplace for specialized resources

To make sharing space and resources easier, The Engine will create an online marketplace for entrepreneurs. A web-based application called the Engine Room will allow entrepreneurs to use or rent specialized resources from each other and from MIT, including office and conference spaces on and off campus, clean rooms, and other facilities and specialized equipment.

The application builds on MIT’s recently released Mobius App, which makes the Institute’s labs, makerspaces, and other resources more broadly available to students. The Engine will work with hosts to establish terms of access, instead of leaving negotiation to each startup. The Engine Room will also serve as a portal to a network of experts and mentors.

In the long run, The Engine aims to link Cambridge, Boston, and other Massachusetts regions and cities as an interconnected network — and to link its activities to those of other centers of innovation across the world, such as MIT’s innovation activities in Hong Kong and Singapore.

Read this article on MIT News.

School of Engineering | MIT News

EECS welcomes six new faculty.

First row (left to right): Adam Belay, Stefanie Mueller, Max Shulaker. Second row (left to right): David Sontag, Ryan Williams, Virginia Vassilevska Williams.

The School of Engineering will welcome 13 new faculty members to its departments, institutes, labs, and centers during the 2016-17 academic year. With research and teaching activities ranging from nuclear fusion to computational complexity theory, they are poised to make vast contributions to new directions across the school and to a range of labs and centers across the Institute.

“We are pleased to welcome such a talented group of faculty to engineering at MIT this year,” says Ian A. Waitz, dean of the School of Engineering. “Every year we broaden the scope and the scale of what we can do, and of how we think about engineering. Our new faculty are often the ones who show us the way forward.”

The six new faculty members joining EECS are:

Adam Belay will join the Department of Electrical Engineering and Computer Science as an assistant professor in July 2017. He holds a PhD in computer science from Stanford University, where he was a member of the secure computer systems group and the multiscale architecture and systems team. Previously, he worked on storage virtualization at VMware Inc. and contributed substantial power-management code to the Linux Kernel project. Belay’s research area is operating systems and networking. Much of his work has focused on restructuring computer systems so that developers can more easily reach the full performance potential of hardware. He received a Stanford graduate fellowship, a VMware graduate fellowship, and an OSDI Jay Lepreau best paper award.

Stefanie Mueller will join the Department of Electrical Engineering and Computer Science as an assistant professor in January 2017. She received her PhD in human-computer interaction (HCI) from the Hasso Plattner Institute in 2016, where she also received an MS in IT-systems engineering. In her research, Mueller develops novel interactive hardware and software systems that advance personal fabrication technologies. Her work has been published at the most selective HCI venues — Association for Computing Machinery (ACM), the Conference for Human Factors in Computing Systems (CHI), and User Interface Software and Technology (UIST) — and received a best paper award and two best-paper nominees. Mueller is an associate chair of the program committees at ACM, CHI, and UIST, and is a general co-chair for the ACM SIGGRAPH Symposium on Computational Fabrication that will take place at MIT in June 2017. She has been an invited speaker at MIT, Stanford, the University of California at Berkeley, Harvard, Carnegie Mellon University, Cornell University, Microsoft Research, Disney Research, Adobe Research, and others. In addition, her work has been covered widely in New Scientist, BBC, The Atlantic, and The Guardian. Mueller will head the HCI engineering group at MIT's Computer Science and Artificial Intelligence Laboratory, which works at the intersection of human-computer interaction, computer graphics, computer vision, and robotics.

Max Shulaker joined the Department of Electrical Engineering and Computer Science as an assistant professor in July. He received his BS, master’s, and PhD in electrical engineering at Stanford, where he was a Fannie and John Hertz Fellow and a Stanford Graduate Fellow. Shulaker’s research focuses on the broad area of nanosystems. His Novel Electronic Systems Group aims to understand and optimize multidisciplinary interactions across the entire computing stack — from low-level synthesis of nanomaterials, to fabrication processes and circuit design for emerging nanotechnologies, up to new architectures — to enable the next generation of high performance and energy-efficient computing systems.

David Sontag will join the Institute for Medical Engineering and Science and the Department of Electrical Engineering and Computer Science in January 2017 as an assistant professor. He earned his bachelor’s degree in computer science at UC Berkeley and his PhD in computer science at MIT, where he worked in Professor Tommi Jaakola’s group on approximate inference and learning in probabilistic models. Sontag is currently an assistant professor of computer science and data science at New York University. Previously, he was a postdoc at Microsoft Research New England. At MIT, his research will focus on machine learning and probabilistic inference, with a particular focus on applications to clinical medicine. He is currently developing algorithms to learn probabilistic models for medical diagnosis directly from unstructured clinical data, automatically discovering and predicting latent (hidden) variables.

Ryan Williams will join MIT as an associate professor (with tenure) in the Department of Electrical Engineering and Computer Science in January 2017, pending the approval of his tenure case by the Executive Committee. He received an BA in computer science and mathematics from Cornell, and a PhD in computer science from Carnegie Mellon. Following postdoctoral appointments at the Institute for Advanced Study (Princeton) and IBM Almaden, he was an assistant professor of computer science at Stanford for five years. Williams’s research interests are in the theoretical design and analysis of efficient algorithms and in computational complexity theory, focusing mainly on new connections (and consequences) forged between algorithm design and logical circuit complexity. Along with some best paper awards, Williams has received a Sloan Fellowship, an NSF CAREER Award, a Microsoft Research Faculty Fellowship, and was an invited speaker at the 2014 International Congress of Mathematicians.

Virginia Vassilevska Williams will join electrical engineering and computer science as an associate professor in January 2017, pending the approval of her case by Academic Council. She received a BS in mathematics and engineering and applied science from Caltech and a PhD in computer science from Carnegie Mellon. She was a postdoctoral fellow at the Institute for Advanced Study (Princeton), UC Berkeley, and Stanford. Prior to joining MIT, she spent three and a half years as an assistant professor at Stanford. Her research interests are broadly in theoretical computer science, focusing on the design and analysis of algorithms and fine-grained complexity. Her work on matrix multiplication algorithms was covered by the press and is the most cited paper in algorithms and complexity in the last five years.

Read about all of the new professors on MIT News.

Larry Hardesty | MIT News

New training technique would reveal the basis for machine-learning systems’ decisions.

Researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) have devised a way to train neural networks so that they provide not only predictions and classifications but rationales for their decisions. Illustration: Christine Daniloff/MIT

In recent years, the best-performing systems in artificial-intelligence research have come courtesy of neural networks, which look for patterns in training data that yield useful predictions or classifications. A neural net might, for instance, be trained to recognize certain objects in digital images or to infer the topics of texts.

But neural nets are black boxes. After training, a network may be very good at classifying data, but even its creators will have no idea why. With visual data, it’s sometimes possible to automate experiments that determine which visual features a neural net is responding to. But text-processing systems tend to be more opaque.

At the Association for Computational Linguistics’ Conference on Empirical Methods in Natural Language Processing, researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) will present a new way to train neural networks so that they provide not only predictions and classifications but rationales for their decisions.

“In real-world applications, sometimes people really want to know why the model makes the predictions it does,” says Tao Lei, an MIT graduate student in electrical engineering and computer science and first author on the new paper. “One major reason that doctors don’t trust machine-learning methods is that there’s no evidence.”

“It’s not only the medical domain,” adds Regina Barzilay, the Delta Electronics Professor of Electrical Engineering and Computer Science and Lei’s thesis advisor. “It’s in any domain where the cost of making the wrong prediction is very high. You need to justify why you did it.”

“There’s a broader aspect to this work, as well,” says Tommi Jaakkola, an MIT professor of electrical engineering and computer science and the third coauthor on the paper. “You may not want to just verify that the model is making the prediction in the right way; you might also want to exert some influence in terms of the types of predictions that it should make. How does a layperson communicate with a complex model that’s trained with algorithms that they know nothing about? They might be able to tell you about the rationale for a particular prediction. In that sense it opens up a different way of communicating with the model.”

Virtual brains

Neural networks are so called because they mimic — approximately — the structure of the brain. They are composed of a large number of processing nodes that, like individual neurons, are capable of only very simple computations but are connected to each other in dense networks.

In a process referred to as “deep learning,” training data is fed to a network’s input nodes, which modify it and feed it to other nodes, which modify it and feed it to still other nodes, and so on. The values stored in the network’s output nodes are then correlated with the classification category that the network is trying to learn — such as the objects in an image, or the topic of an essay.

Over the course of the network’s training, the operations performed by the individual nodes are continuously modified to yield consistently good results across the whole set of training examples. By the end of the process, the computer scientists who programmed the network often have no idea what the nodes’ settings are. Even if they do, it can be very hard to translate that low-level information back into an intelligible description of the system’s decision-making process.

In the new paper, Lei, Barzilay, and Jaakkola specifically address neural nets trained on textual data. To enable interpretation of a neural net’s decisions, the CSAIL researchers divide the net into two modules. The first module extracts segments of text from the training data, and the segments are scored according to their length and their coherence: The shorter the segment, and the more of it that is drawn from strings of consecutive words, the higher its score.

The segments selected by the first module are then passed to the second module, which performs the prediction or classification task. The modules are trained together, and the goal of training is to maximize both the score of the extracted segments and the accuracy of prediction or classification.

One of the data sets on which the researchers tested their system is a group of reviews from a website where users evaluate different beers. The data set includes the raw text of the reviews and the corresponding ratings, using a five-star system, on each of three attributes: aroma, palate, and appearance.

What makes the data attractive to natural-language-processing researchers is that it’s also been annotated by hand, to indicate which sentences in the reviews correspond to which scores. For example, a review might consist of eight or nine sentences, and the annotator might have highlighted those that refer to the beer’s “tan-colored head about half an inch thick,” “signature Guinness smells,” and “lack of carbonation.” Each sentence is correlated with a different attribute rating.

Validation

As such, the data set provides an excellent test of the CSAIL researchers’ system. If the first module has extracted those three phrases, and the second module has correlated them with the correct ratings, then the system has identified the same basis for judgment that the human annotator did.

In experiments, the system’s agreement with the human annotations was 96 percent and 95 percent, respectively, for ratings of appearance and aroma, and 80 percent for the more nebulous concept of palate.

In the paper, the researchers also report testing their system on a database of free-form technical questions and answers, where the task is to determine whether a given question has been answered previously.

In unpublished work, they’ve applied it to thousands of pathology reports on breast biopsies, where it has learned to extract text explaining the bases for the pathologists’ diagnoses. They’re even using it to analyze mammograms, where the first module extracts sections of images rather than segments of text.

“There’s a lot of hype now — and rightly so — around deep learning, and specifically deep learning for natural-language processing,” says Byron Wallace, an assistant professor of computer and information science at Northeastern University. “But a big drawback for these models is that they’re often black boxes. Having a model that not only makes very accurate predictions but can also tell you why it’s making those predictions is a really important aim.”

“As it happens, we have a paper that’s similar in spirit being presented at the same conference,” Wallace adds. “I didn’t know at the time that Regina was working on this, and I actually think hers is better. In our approach, during the training process, while someone is telling us, for example, that a movie review is very positive, we assume that they’ll mark a sentence that gives you the rationale. In this way we train the deep-learning model to extract these rationales. But they don’t make this assumption, so their model works without using direct annotations with rationales, which is a very nice property.”

Read this article on MIT News.

The Department of Electrical Engineering and Computer Science is proud to present the following awards for SM and PhD theses for Fall 2016. 

2016 George M. Sprowls Award for outstanding PhD thesis in CS
Haitham Al Hassanieh
The Sparse Fourier Transform: Theory and Practice
Supervisor: Dina Katabi

Haogang Chen
Certifying a Crash-safe File System
Supervisor: Frans Kaashoek and Nickolai Zeldovich

Myers A. Davis
Visual Vibration Analysis
Supervisor: Fredo Durand

Christopher W. Fletcher
Oblivious RAM: From Theory to Practice
Supervisor: Srinivas Devadas

Yin Tat Lee
Faster Algorithms for Convex and Combinatorial Optimization
Supervisor: Jonathan Kelner

2016 Jin-Au Kong Award for outstanding PhD thesis in EE
Adam McCaughan
Superconducting Thin Film Nanoelectrics
Supervisor: Karl K. Berggren

Rebecca Mieloszyk
Learning and Model-Based Approaches to Improved Patient Monitoring, Assessment and Treatment in Capnography and Procedual Sedation
Supervisors: Thomas Heldt and George Verghese

2015 Ernst A. Guillemin Thesis Award for outstanding SM thesis in EE
Wei Ouyang
Microfluidic Platform for Rapid Biologics Activity Assessment using Molecular Charge Modulation and Electrokinetic Concentration based Receptor Assays
Supervisor: Jongyoon Han

Clément Pit Claudel
Compilation using Correct-by-Construction Program Synthesis
Supervisor: Adam Chlipala

2015 William A. Martin Memorial Thesis Award for outstanding thesis in CS
Guowei Zhang
Architectural Support to Exploit Commutativity in Shared-Memory Systems
Supervisor: Daniel Sanchez

Larry Hardesty | MIT News

Self-driving scooter demonstrated at MIT complements autonomous golf carts and city cars.

An autonomous mobility scooter and related software were designed by researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL), the National University of Singapore, and the Singapore-MIT Alliance for Research and Technology (SMART). Courtesy of the Autonomous Vehicle Team of the SMART Future of Urban Mobility Project

At MIT’s 2016 Open House last spring, more than 100 visitors took rides on an autonomous mobility scooter in a trial of software designed by researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL), the National University of Singapore, and the Singapore-MIT Alliance for Research and Technology (SMART).

The researchers had previously used the same sensor configuration and software in trials of autonomous cars and golf carts, so the new trial completes the demonstration of a comprehensive autonomous mobility system. A mobility-impaired user could, in principle, use a scooter to get down the hall and through the lobby of an apartment building, take a golf cart across the building’s parking lot, and pick up an autonomous car on the public roads.

The new trial establishes that the researchers’ control algorithms work indoors as well as out. “We were testing them in tighter spaces,” says Scott Pendleton, a graduate student in mechanical engineering at the National University of Singapore (NUS) and a research fellow at SMART. “One of the spaces that we tested in was the Infinite Corridor of MIT, which is a very difficult localization problem, being a long corridor without very many distinctive features. You can lose your place along the corridor. But our algorithms proved to work very well in this new environment.”

The researchers’ system includes several layers of software: low-level control algorithms that enable a vehicle to respond immediately to changes in its environment, such as a pedestrian darting across its path; route-planning algorithms; localization algorithms that the vehicle uses to determine its location on a map; map-building algorithms that it uses to construct the map in the first place; a scheduling algorithm that allocates fleet resources; and an online booking system that allows users to schedule rides.

Uniformity

Using the same control algorithms for all types of vehicles — scooters, golf carts, and city cars — has several advantages. One is that it becomes much more practical to perform reliable analyses of the system’s overall performance.

“If you have a uniform system where all the algorithms are the same, the complexity is much lower than if you have a heterogeneous system where each vehicle does something different,” says Daniela Rus, the Andrew and Erna Viterbi Professor of Electrical Engineering and Computer Science at MIT and one of the project’s leaders. “That’s useful for verifying that this multilayer complexity is correct.”

Furthermore, with software uniformity, information that one vehicle acquires can easily be transferred to another. Before the scooter was shipped to MIT, for instance, it was tested in Singapore, where it used maps that had been created by the autonomous golf cart.

Similarly, says Marcelo Ang, an associate professor of mechanical engineering at NUS who co-leads the project with Rus, in ongoing work the researchers are equipping their vehicles with machine-learning systems, so that interactions with the environment will improve the performance of their navigation and control algorithms. “Once you have a better driver, you can easily transplant that to another vehicle,” says Ang. “That’s the same across different platforms.”

Finally, software uniformity means that the scheduling algorithm has more flexibility in its allocation of system resources. If an autonomous golf cart isn’t available to take a user across a public park, a scooter could fill in; if a city car isn’t available for a short trip on back roads, a golf cart might be. “I can see its usefulness in large indoor shopping malls and amusement parks to take [mobility-impaired] people from one spot to another,” says Dan Ding, an associate professor of rehabilitation science and technology at the University of Pittsburgh, about the system.

Changing perceptions

The scooter trial at MIT also demonstrated the ease with which the researchers could deploy their modular hardware and software system in a new context. “It’s extraordinary to me, because it’s a project that the team conducted in about two months,” Rus says. MIT’s Open House was at the end of April, and “the scooter didn’t exist on February 1st,” Rus says.

The researchers described the design of the scooter system and the results of the trial in a paper they presented last week at the IEEE International Conference on Intelligent Transportation Systems. Joining Rus, Pendleton, and Ang on the paper are You Hong Eng, who leads the SMART autonomous-vehicle project, and four other researchers from both NUS and SMART.

The paper also reports the results of a short user survey that the researchers conducted during the trial. Before riding the scooter, users were asked how safe they considered autonomous vehicles to be, on a scale from one to five; after their rides, they were asked the same question again. Experience with the scooter brought the average safety score up, from 3.5 to 4.6.

Read this article on MIT News.

Larry Hardesty | MIT News

New system lets nonexperts optimize programs that run on multiprocessor chips.

A system developed by researchers at MIT and Stony Brook University should make it easier for researchers to solve complex computational problems using dynamic programming optimized for multicore chips — without the expertise that such programming typically requires. Image: MIT News (figures courtesy of the researchers)

Dynamic programming is a technique that can yield relatively efficient solutions to computational problems in economics, genomic analysis, and other fields. But adapting it to computer chips with multiple “cores,” or processing units, requires a level of programming expertise that few economists and biologists have.

Researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and Stony Brook University aim to change that, with a new system that allows users to describe what they want their programs to do in very general terms. It then automatically produces versions of those programs that are optimized to run on multicore chips. It also guarantees that the new versions will yield exactly the same results that the single-core versions would, albeit much faster.

In experiments, the researchers used the system to “parallelize” several algorithms that used dynamic programming, splitting them up so that they would run on multicore chips. The resulting programs were between three and 11 times as fast as those produced by earlier techniques for automatic parallelization, and they were generally as efficient as those that were hand-parallelized by computer scientists.

The researchers presented their new system last week at the Association for Computing Machinery’s conference on Systems, Programming, Languages and Applications: Software for Humanity.

Dynamic programming offers exponential speedups on a certain class of problems because it stores and reuses the results of computations, rather than recomputing them every time they’re required.

“But you need more memory, because you store the results of intermediate computations,” says Shachar Itzhaky, first author on the new paper and a postdoc in the group of Armando Solar-Lezama, an associate professor of electrical engineering and computer science at MIT. “When you come to implement it, you realize that you don't get as much speedup as you thought you would, because the memory is slow. When you store and fetch, of course, it’s still faster than redoing the computation, but it’s not as fast as it could have been.”

Outsourcing complexity

Computer scientists avoid this problem by reordering computations so that those requiring a particular stored value are executed in sequence, minimizing the number of times that the value has to be recalled from memory. That’s relatively easy to do with a single-core computer, but with multicore computers, when multiple cores are sharing data stored at multiple locations, memory management become much more complex. A hand-optimized, parallel version of a dynamic-programming algorithm is typically 10 times as long as the single-core version, and the individual lines of code are more complex, to boot.

The CSAIL researchers’ new system — dubbed Bellmania, after Richard Bellman, the applied mathematician who pioneered dynamic programming — adopts a parallelization strategy called recursive divide-and-conquer. Suppose that the task of a parallel algorithm is to perform a sequence of computations on a grid of numbers, known as a matrix. Its first task might be to divide the grid into four parts, each to be processed separately.

But then it might divide each of those four parts into four parts, and each of those into another four parts, and so on. Because this approach — recursion — involves breaking a problem into smaller subproblems, it naturally lends itself to parallelization.

Joining Itzhaky on the new paper are Solar-Lezama; Charles Leiserson, the Edwin Sibley Webster Professor of Electrical Engineering and Computer Science; Rohit Singh and Kuat Yessenov, who were MIT both graduate students in electrical engineering and computer science when the work was done; Yongquan Lu, an MIT undergraduate who participated in the project through MIT’s Undergraduate Research Opportunities Program; and Rezaul Chowdhury, an assistant professor of computer science at Stony Brook, who was formerly a research affiliate in Leiserson’s group.

Leiserson’s group specializes in divide-and-conquer parallelization techniques; Solar-Lezama’s specializes in program synthesis, or automatically generating code from high-level specifications. With Bellmania, the user simply has to describe the first step of the process — the division of the matrix and the procedures to be applied to the resulting segments. Bellmania then determines how to continue subdividing the problem so as to use memory efficiently.

Rapid search

At each level of recursion — with each successively smaller subdivision of the matrix — a program generated by Bellmania will typically perform some operation on some segment of the matrix and farm the rest out to subroutines, which can be performed in parallel. Each of those subroutines, in turn, will perform some operation on some segment of the data and farm the rest out to further subroutines, and so on.

Bellmania determines how much data should be processed at each level and which subroutines should handle the rest. “The goal is to arrange the memory accesses such that when you read a cell [of the matrix], you do as much computation as you can with it, so that you will not have to read it again later,” Itzhaky says.

Finding the optimal division of tasks requires canvassing a wide range of possibilities. Solar-Lezama’s group has developed a suite of tools to make that type of search more efficient; even so, Bellmania takes about 15 minutes to parallelize a typical dynamic-programming algorithm. That’s still much faster than a human programmer could perform the same task, however. And the result is guaranteed to be correct; hand-optimized code is so complex that it’s easy for errors to creep in.

“The work that they’re doing is really foundational in enabling a broad set of applications to run on multicore and parallel processors,” says David Bader, a professor of computational science and engineering at Georgia Tech. “One challenge has been to enable high-level writing of programs that work on our current multicore processors, and up to now doing that requires heroic, low-level manual coding to get performance. What they provide is a much simpler, high-level technique for some classes of programs that makes it very easy to write the program and have their system automatically figure out how to divide up the work to create codes that are competitive with hand-tuned, low-level coding.

“The types of applications that they would enable range from computational biology, to proteomics, to cybersecurity, to sorting, to scheduling problems of all sorts, to managing network traffic — there are countless examples of real algorithms in the real world for which they now enable much more efficient code,” he adds. “It’s remarkable.”

Read this article on MIT News.

Terri Park | MIT Innovation Institute

Nanoscience pioneer and Institute Professor Millie Dresselhaus accepted the medal in Paris on behalf of the Institute.

Institute Professor Millie Dresselhaus accepts the UNESCO Medal from UNESCO Director-General Irina Bokova at the awards ceremony on Oct. 11 in Paris, France. Photo: I. Marin/UNESCO

MIT has been honored with the UNESCO Medal for contributions to the development of nanoscience and nanotechnologies by the United Nations Educational, Scientific and Cultural Organization (UNESCO).

Established in 2010, the UNESCO Medal has awarded over 30 prominent scientists and public figures for their individual contributions to advancing the fields of nanoscience and nanotechnologies. This year MIT shares the distinction, along with St. Petersburg State University of Information Technologies in Russia, of being the first organization to be recognized. In addition to the two universities, four eminent scientists from Korea, the United Arab Emirates, Ukraine, and the United Kingdom, were recipients of the medal.

An awards ceremony was held on Oct. 11 at the UNESCO headquarters in Paris, France. Institute Professor Mildred (Millie) Dresselhaus, a nanoscience pioneer who herself has won many recognitions including the U.S. Presidential Medal of Freedom and the L'Oreal-UNESCO Award for Women in Science, made the trip at the invitation of President Rafael Reif to accept the medal on behalf of MIT.

“Using science and technology as a way to bring people together is something MIT has learned to do really well,” says Dresselhaus. “Our faculty, staff, and students come together from countries all over the world with diverse technical backgrounds to work across the many academic departments and laboratories on campus. This culture of interdisciplinary collaboration enables us to work for common goals, so it made sense to me that MIT was recognized as an institution. This should serve as encouragement to move forward as rapidly as possible to complete MIT.nano and to achieve some exceptionally great outcomes through this initiative as it comes to fruition.”

The award will eventually be displayed within the public spaces of MIT.nano — the 214,000-square-foot center for nanoscience and nanotechnology that is currently under construction in the heart of the MIT campus — after the building opening in June 2018, says Vladimir Bulović, faculty lead of the project.

The UNESCO Medal is an initiative of the International Commission responsible for developing the Encyclopedia of Life Support Systems theme on nanoscience and nanotechnologies. Each year, the medal recognizes those making significant contributions in the field in an effort to showcase the tremendous benefits of progress being made. MIT joins a distinguished group of scientists who have received the medal thus far, including Nobel Prize-winners in physics Zhores Alferov and Isamu Akasaki.

Read this article on MIT News.

Larry Hardesty | MIT News

“Information extraction” system helps turn plain text into data for statistical analysis.

Information extraction — or automatically classifying data items stored as plain text — is a major topic of artificial-intelligence research. Image: MIT News

Of the vast wealth of information unlocked by the Internet, most is plain text. The data necessary to answer myriad questions — about, say, the correlations between the industrial use of certain chemicals and incidents of disease, or between patterns of news coverage and voter-poll results — may all be online. But extracting it from plain text and organizing it for quantitative analysis may be prohibitively time consuming.

Information extraction — or automatically classifying data items stored as plain text — is thus a major topic of artificial-intelligence research. Last week, at the Association for Computational Linguistics’ Conference on Empirical Methods on Natural Language Processing, researchers from MIT’s Computer Science and Artificial Intelligence Laboratory won a best-paper award for a new approach to information extraction that turns conventional machine learning on its head.

Most machine-learning systems work by combing through training examples and looking for patterns that correspond to classifications provided by human annotators. For instance, humans might label parts of speech in a set of texts, and the machine-learning system will try to identify patterns that resolve ambiguities — for instance, when “her” is a direct object and when it’s an adjective.

Typically, computer scientists will try to feed their machine-learning systems as much training data as possible. That generally increases the chances that a system will be able to handle difficult problems.

In their new paper, by contrast, the MIT researchers train their system on scanty data — because in the scenario they’re investigating, that’s usually all that’s available. But then they find the limited information an easy problem to solve.

“In information extraction, traditionally, in natural-language processing, you are given an article and you need to do whatever it takes to extract correctly from this article,” says Regina Barzilay, the Delta Electronics Professor of Electrical Engineering and Computer Science and senior author on the new paper. “That’s very different from what you or I would do. When you’re reading an article that you can’t understand, you’re going to go on the web and find one that you can understand.”

Confidence boost

Essentially, the researchers’ new system does the same thing. A machine-learning system will generally assign each of its classifications a confidence score, which is a measure of the statistical likelihood that the classification is correct, given the patterns discerned in the training data. With the researchers’ new system, if the confidence score is too low, the system automatically generates a web search query designed to pull up texts likely to contain the data it’s trying to extract.

It then attempts to extract the relevant data from one of the new texts and reconciles the results with those of its initial extraction. If the confidence score remains too low, it moves on to the next text pulled up by the search string, and so on.

“The base extractor isn’t changing,” says Adam Yala, a graduate student in the MIT Department of Electrical Engineering and Computer Science (EECS) and one of the coauthors on the new paper. “You’re going to find articles that are easier for that extractor to understand. So you have something that’s a very weak extractor, and you just find data that fits it automatically from the web.” Joining Yala and Barzilay on the paper is first author Karthik Narasimhan, also a graduate student in EECS.

Remarkably, every decision the system makes is the result of machine learning. The system learns how to generate search queries, gauge the likelihood that a new text is relevant to its extraction task, and determine the best strategy for fusing the results of multiple attempts at extraction.

Just the facts

In experiments, the researchers applied their system to two extraction tasks. One was the collection of data on mass shootings in the U.S., which is an essential resource for any epidemiological study of the effects of gun-control measures. The other was the collection of similar data on instances of food contamination. The system was trained separately for each task.

In the first case — the database of mass shootings — the system was asked to extract the name of the shooter, the location of the shooting, the number of people wounded, and the number of people killed. In the food-contamination case, it extracted food type, type of contaminant, and location. In each case, the system was trained on about 300 documents.

From those documents, it learned clusters of search terms that tended to be associated with the data items it was trying to extract. For instance, the names of mass shooters were correlated with terms like “police,” “identified,” “arrested,” and “charged.” During training, for each article the system was asked to analyze, it pulled up, on average, another nine or 10 news articles from the web.

The researchers compared their system’s performance to that of several extractors trained using more conventional machine-learning techniques. For every data item extracted in both tasks, the new system outperformed its predecessors, usually by about 10 percent.

“One of the difficulties of natural language is that you can express the same information in many, many different ways, and capturing all that variation is one of the challenges of building a comprehensive model,” says Chris Callison-Burch, an assistant professor of computer and information science at the University of Pennsylvania. “[Barzilay and her colleagues] have this super-clever part of the model that goes out and queries for more information that might result in something that’s simpler for it to process. It’s clever and well-executed.”

Callison-Burch’s group is using a combination of natural-language processing and human review to build a database of information on gun violence, much like the one that the MIT researchers’ system was trained to produce. “We’ve crawled millions and millions of news articles, and then we pick out ones that the text classifier thinks are related to gun violence, and then we have humans start doing information extraction manually,” he says. “Having a model like Regina’s that would allow us to predict whether or not this article corresponded to one that we’ve already annotated would be a huge time savings. It’s something that I’d be very excited to do in the future.”

Read this article on MIT News.