Imaging system helps surgeons remove tiny ovarian tumors

Ovarian cancer is usually diagnosed only after it has reached an advanced stage, with many tumors spread throughout the abdomen. Most patients undergo surgery to remove as many of these tumors as possible, but because some are so small and widespread, it is difficult to eradicate all of them.

Researchers at MIT, working with surgeons and oncologists at Massachusetts General Hospital (MGH), have now developed a way to improve the accuracy of this surgery, called debulking. Using a novel fluorescence imaging system, they were able to find and remove tumors as small as 0.3 millimeters — smaller than a poppy seed — during surgery in mice. Mice that underwent this type of image-guided surgery survived 40 percent longer than those who had tumors removed without the guided system.

“What’s nice about this system is that it allows for real-time information about the size, depth, and distribution of tumors,” says Angela Belcher, the James Mason Crafts Professor of Biological Engineering and Materials Science at MIT, a member of the Koch Institute for Integrative Cancer Research, and the recently appointed head of MIT’s Department of Biological Engineering.

The researchers are now seeking FDA approval for a phase 1 clinical trial to test the imaging system in human patients. In the future, they hope to adapt the system for monitoring patients at risk for tumor recurrence, and eventually for early diagnosis of ovarian cancer, which is easier to treat if it is caught earlier.

Belcher and Michael Birrer, formerly the director of medical gynecologic oncology at MGH and now the director of the O’Neal Comprehensive Cancer Center at the University of Alabama at Birmingham, are the senior authors of the study, published online in the journal ACS Nano this week.

Neelkanth Bardhan, a Mazumdar-Shaw International Oncology Fellow at the Koch Institute, and Lorenzo Ceppi, a researcher at MGH, are the lead authors of the paper. Other authors include MGH researcher YoungJeong Na, MIT Lincoln Laboratory technical staff members Andrew Siegel and Nandini Rajan, Robert Fruscio of the University of Milan-Bicocca, and Marcela del Carmen, a gynecologic oncologist at MGH and chief medical officer of the Massachusetts General Physicians Organization.

Glowing tumors

Because there is no good way to detect early-stage ovarian cancer, it is one of the most difficult types of cancer to treat. Of 250,000 new cases diagnosed each year worldwide, 75 percent are in an advanced stage. In the United States, the five-year combined survival rate for all stages of ovarian cancer is 47 percent, only a slight improvement from 38 percent three decades ago, despite the advent of chemotherapeutic drugs such as cisplatin, approved by the FDA in 1978 for ovarian cancer treatment. In contrast, the five-year combined survival rate for all stages of breast cancer has steadily improved, from around 75 percent in the 1970s to over 90 percent now.  

“We desperately need better upfront therapies, including surgery, for these (ovarian cancer) patients,” Birrer says.

Belcher and Birrer joined forces to work on this problem through the Bridge Project, a collaboration between the Koch Institute and Dana-Farber/Harvard Cancer Center. Belcher’s lab has been developing a novel type of medical imaging based on light in the near-infrared (NIR) spectrum. In a paper published in March, she reported that this imaging system could achieve an unprecedented combination of resolution and penetration-depth in living tissue.

In the new study, Belcher, Birrer, and their colleagues worked with researchers at MIT Lincoln Laboratory to adapt NIR imaging to help surgeons locate tumors during ovarian cancer surgery, by providing continuous, real-time imaging of the abdomen, with tumors highlighted by fluorescence. Previous analyses have shown that survival rates are strongly inversely correlated with the amount of residual tumor mass left behind in the patient during debulking surgery, but many ovarian tumors are so small or hidden that surgeons can’t find them.

To make the tumors visible, the researchers designed chemical probes using single-walled carbon nanotubes that emit fluorescent light when illuminated by a laser. They coated these nanotubes with a peptide that binds to SPARC, a protein that is overexpressed by highly invasive ovarian cancer cells. This probe binds to the tumors and makes them fluoresce at NIR wavelengths, allowing surgeons to more easily find them with fluorescence imaging.

The researchers tested the image-guided system in mice that had ovarian tumors implanted in a region of the abdominal cavity known as the intraperitoneal space, and showed that surgeons were able to locate and remove tumors as small as 0.3 millimeters. Ten days after surgery, these mice had no detectable tumors, while mice that had undergone the traditional, non-image-guided surgery, had many residual tumors missed by the surgeon.

By three weeks after the surgery, many of the tumors had grown back in the mice that underwent image-guided surgery, but those mice still had a median survival rate that was 40 percent longer than that of mice that underwent traditional surgery.

No other imaging system would be able to locate tumors that small during a surgical procedure, the researchers say.

“You can’t have a patient in a CT machine or an MRI machine and have the surgeon perform this surgical debulking procedure at the same time, and you can’t expose the patient to X-ray radiation for multiple hours of the long surgery. This optics-based imaging system allows us to do that in a safe manner,” Bardhan says.

Alessandro Santin, a professor of obstetrics and gynecology and clinical research program leader at the Yale University School of Medicine, described the results as “intriguing.”

“These data support the potential use of this novel imaging system in the intraoperative setting for the optical detection of residual malignant tissue at the time of surgical staging, and/or cytoreductive surgery in ovarian cancer patients,” says Santin, who was not involved in the study.

Monitoring patients

For most ovarian cancer patients, tumor debulking surgery is followed by chemotherapy, so the researchers now plan to do another study where they treat the mice with chemotherapy after image-guided surgery, in hopes of preventing the remaining tiny tumors from spreading.

“We know that the amount of tumor removed at the time of surgery for patients with advanced-stage ovarian cancer is directly correlated with their outcome,” Birrer says. “This imaging device will now allow the surgeon to go beyond the limits of resecting tumors visible to the naked eye, and should usher in a new age of effective debulking surgery.”

Now that they have demonstrated that this concept can be successfully applied to imaging during surgery, the researchers hope to begin adapting the system for use in human patients.

“In principle, it’s quite doable,” Siegel says. “It’s purely the mechanics and the funding at this point, because this mouse experiment serves as the proof of principle and may actually have been more challenging than building a human-scale system.”

The researchers also hope to deploy this type of imaging to monitor patients after surgery, and eventually to develop it as a diagnostic tool for screening women at high risk for developing ovarian cancer.

“A major focus for us right now is developing the technology to be able diagnose ovarian cancer early, in stage 1 or stage 2, before the disease becomes disseminated,” Belcher says. “That could have a huge impact on survival rates, because survival is related to the stage of detection.”

The research was funded, in part, by the Bridge Project and the Koch Institute Support (core) Grant from the National Cancer Institute, with previous support for the development of the system from the Koch Institute Frontier Research Program and the Kathy and Curt Marble Cancer Research Fund.

de MIT News http://bit.ly/2vs6YUR

Water Innovation Prize goes to startups targeting methane and wastewater

A startup with a cheap technology for purifying textile wastewater and another with a system to help reduce methane emissions from cattle were named co-winners of the MIT Water Innovation Prize on Thursday.

After eight student finalist teams pitched their companies’ water-related solutions, the judges couldn’t agree on the winners and ultimately split the grand prize into two $14,000 checks for the co-winners.

The founders of both the seaweed-producing startup Symbrosia and the textile wastewater purification startup SiPure said they were happy to split the winnings.

“We were just so proud to be here,” SiPure business development lead Lily Cheng Zedler said after the event. “We’re really grateful to the Water Innovation Prize and the judges for believing in us.”

Close to 200 people, including students, faculty, investors, and people working in the private industry, traveled to the sixth floor of the Media Lab for the event. Members from the eight finalist teams came from as far away as Lebanon and as close as the MIT Sloan School of Management to share their ideas.

The third place, $7,000 prize went to Volta Irrigation, which loans seeds, fertilizers, and pesticides to smallholder farmers in Rwanda and surrounding countries, then helps the farmers increase their productivity by loaning them a proprietary irrigation system called the Alma Volta. The stationary, bicycle-like device works by having operators pedal, which powers an inverter, battery, and pump that efficiently distribute over 3,000 liters of water per hour onto crops.

Addressing livestock methane emissions

According to the Environmental Protection Agency, methane accounts for about 10 percent of U.S. greenhouse gas emissions. The largest source of methane is livestock such as cows, pigs, and goats, who produce it as part of their normal digestive processes.

Recent research has shown that mixing just 2 percent of a specific kind of algae into a cow’s diet can reduce their methane emissions by 99 percent.

Symbrosia is acting on those findings with a patent-pending system that consists of a tank for growing that algae, a tank for growing shrimp, and chambers that move waste and water back and forth. When waste from the shrimp moves to the algae tank it acts as fertilizer, and as the algae absorb nutrients from the water it produces clean, oxygenated water for the shrimp.

The result is a weekly harvest of algae and local, organic shrimp (which are grown in three-month rotational cycles). The only water loss in the system is due to evaporation, and all of the waste is dissolved back into the water, according to the company. Symbrosia plans to sell the algae to feed suppliers at $1.60 a pound, and the shrimp to restaurants at $24 a pound.

With its algae, the company plans to first target the mixed-ration dairy feed supplement market, estimated to be around $5.3 billion in size. With its shrimp, the company will first target the $31 million imported organic shrimp market in the U.S.

The company will begin its first pilot project with three corporate partners at Port Hueneme in California toward the end of this year. Eventually, it plans to place large versions of its system near livestock industry hot spots to maximize its impact.

Cleaning up the textile industry

Garment manufacturers use huge volumes of water each year to dye fabrics. Purifying the resulting wastewater is a complex, expensive process that can account for up to 25 percent of the operating costs of a standard textile mill.

Unfortunately, the low margins in the textile industry lead many manufacturers to dump the wastewater in local waterways. For example, in India, the world’s second largest producer of textiles, 80 percent of textile wastewater goes untreated, according to SiPure.

Wastewater dumping leads to the pollution of drinking water, destruction of local agriculture, and long-term health consequences for people in the area.

SiPure has developed and patented a silicon membrane that it says makes the process of purifying textile wastewater dramatically simpler and cheaper. Billions of tiny nanopores within the membrane allow water to flow through while molecular dyes get stuck.

“It looks boring on the surface, just a gray square,” SiPure co-founder Brendan Smith, who invented the technology during his PhD work in MIT’s Department of Materials Science and Engineering, told the audience during the pitch. “But the magic is in the cross section.”

Smith says the membrane is capable of removing more than 99 percent of the dyes in waters and can be produced for about 10 times less than competing ceramic-based purification technology. SiPure says its membrane also decreases maintenance costs while working for around 10 years.

This summer, the company is starting a pilot project with a textile mill in India, where its membranes will purify 50 to 100 liters of wastewater each day. From there, the founders plan to continue scaling throughout India in hopes of capturing 35 to 40 percent of the market by 2025.

The Water Innovation Prize, which helps translate research and ideas into business and impact, has been hosted by the MIT Water Club since 2015. Each year, student-led finalist teams from around the country and, increasingly, the world, come to MIT’s campus to pitch their water-related innovations.

de MIT News http://bit.ly/2IDTEoF

“Nanofiber yarn” makes for stretchy, protective artificial tissue

The human body is held together by an intricate cable system of tendons and muscles, engineered by nature to be tough and highly stretchable. An injury to any of these tissues, particularly in a major joint like the shoulder or knee, can require surgical repairs and weeks of limited mobility to fully heal.

Now MIT engineers have come up with a tissue engineering design that may enable flexible range of motion in injured tendons and muscles during healing.

The team has engineered small coils lined with living cells, that they say could act as stretchy scaffolds for repairing damaged muscles and tendons. The coils are made from hundreds of thousands of biocompatible nanofibers, tightly twisted into coils resembling miniature nautical rope, or yarn.

The researchers coated the yarn with living cells, including muscle and mesenchymal stem cells, which naturally grow and align along the yarn, into patterns similar to muscle tissue. The researchers found the yarn’s coiled configuration helps to keep cells alive and growing, even as the team stretched and bent the yarn multiple times.

In the future, the researchers envision doctors could line patients’ damaged tendons and muscles with this new flexible material, which would be coated with the same cells that make up the injured tissue. The “yarn’s” stretchiness could help maintain a patient’s range of motion while new cells continue to grow to replace the injured tissue.

“When you repair muscle or tendon, you really have to fix their movement for a period of time, by wearing a boot, for example,” says Ming Guo, assistant professor of mechanical engineering at MIT. “With this nanofiber yarn, the hope is, you won’t have to wearing anything like that.”

Guo and his colleagues published their results this week in the Proceedings of the National Academy of Sciences. His MIT co-authors are Yiwei Li, Yukun Hao, Satish Gupta, and Jiliang Hu. The team also includes Fengyun Guo, Yaqiong Wang, Nü Wang, and Yong Zhao, of Beihang University.

Stuck on gum

The new nanofiber yarn was inspired in part by the group’s previous work on lobster membranes, where they found the crustacean’s tough yet stretchy underbelly is due to a layered, plywood-like structure. Each microscopic layer contains hundreds of thousands of nanofibers, all aligned in the same direction, at an angle that is slightly offset from the layer just above and below.

The nanofibers’ precise alignment makes each individual layer highly stretchable in the direction in which the fibers are arranged. Guo, whose work focuses on biomechanics, saw the lobster’s natural stretchy patterning as an inspiration for designing artificial tissues, particularly for high-stretch regions of the body such as the shoulder and knee.

Guo says biomedical engineers have embedded muscle cells in other stretchy materials such as hydrogels, in attempts to fashion flexible artificial tissues. However, while the hydrogels themselves are stretchy and tough, the embedded cells tend to snap when stretched, like tissue paper stuck on a piece of gum.

“When you largely deform a material like hydrogel, it will be stretched just fine, but the cells can’t take it,” Guo says. “A living cell is sensitive, and when you stretch them, they die.”

Shelter in a slinky

The researchers realized that simply considering the stretchability of a material would not be enough to design an artificial tissue. That material would also have to be able to protect cells from the severe strains produced when the material is stretched.

The team looked to actual muscles and tendons for further inspiration, and observed that the tissues are made from strands of aligned protein fibers, coiled together to form microscopic helices, along which muscle cells grow. It turns out that, when the protein coils stretch out, the muscle cells simply rotate, like tiny pieces of tissue paper stuck on a slinky.

Guo looked to replicate this natural, stretchy, cell-protecting structure as an artificial tissue material. To do so, the team first created hundreds of thousands of aligned nanofibers, using electrospinning, a technique that uses electric force to spin ultrathin fibers out from a solution of polymer or other materials. In this case, he generated nanofibers made from biocompatible materials such as cellulose.

The team then bundled aligned fibers together and twisted them slowly to form first a spiral, and then an even tighter coil, ultimately resembling yarn and measuring about half a millimeter wide. Finally, they seeded live cells along each coil, including muscle cells, mesenchymal stem cells, and human breast cancer cells.

The researchers then repeatedly stretched each coil up to six times its original length, and found that the majority of cells on each coil remained alive and continued to grow as the coils were stretched. Interestingly, when they seeded cells on looser, spiral-shaped structures made from the same materials, they found cells were less likely to remain alive. Guo says the structure of the tighter coils seems to “shelter” cells from damage.

Going forward, the group plans to fabricate similar coils from other biocompatible materials such as silk, which could ultimately be injected into an injured tissue. The coils could provide a temporary, flexible scaffold for new cells to grow. Once the cells successfully repair an injury, the scaffold can dissolve away.

“We may be able to one day embed these structures under the skin, and the [coil] material would eventually be digested, while the new cells stay put,” Guo says. “The nice thing about this method is, it’s really general, and we can try different materials. This may push the limit of tissue engineering a lot.”

This research was funded, in part, by MIT Research Support Committee Fund.

de MIT News http://bit.ly/2Ve5LPt

NEET welcomes Boeing sponsorship

A cross-departmental engineering program focused on modern industry and real-world projects is welcoming a new sponsor and industry collaborator: aerospace company Boeing.

This is a deeply beneficial industry collaboration for the New Engineering Education Transformation (NEET) program, which launched in 2017 to reimagine engineering education at MIT, says NEET Executive Director Amitava "Babi" Mitra.

A cross-disciplinary endeavor with a focus on integrative, project-centric learning, NEET cultivates the essential skills, knowledge, and qualities to help students build the “new machines and systems” that will be required to address the formidable challenges posed by the 21st century.

“Industry experts bring a flavor of real life into the projects,” Mitra says of the NEET program, which is offering five threads — Advanced Materials Machines, Autonomous Machines, Digital Cities, Living Machines, and Renewable Energy Machines.

“Projects are more engaging for students when we can connect them to what’s happening in industry,” says Mitra. Students in NEET earn a degree in their chosen major and are simultaneously awarded a NEET certificate in their chosen thread.

This fall, Boeing will become a founding co-sponsor of the Autonomous Machines thread. Alexa Jan, a senior in the Department of Electrical Engineering and Computer Science and NEET participant, calls this news exciting.

“A collaboration with Boeing will provide more resources for interdisciplinary projects that will help students be leaders in the autonomous machines field and beyond,” she says.

The close proximity of Aurora Flight Sciences, a Boeing company, will enable frequent interactions between students and autonomy and robotics experts from industry, says Jonathan P. How, the Richard Cockburn Maclaurin Professor of Aeronautics and Astronautics, who leads the Autonomous Machines thread.

“This collaboration will provide students with a wealth of experience on the needs for advanced autonomy in complex systems and a better understanding of how to implement those algorithms,” says How. He has worked with Boeing on research for more than a decade.

The collaboration was celebrated when Boeing representatives visited campus visit on April 17 for a series of meetings with administrators, faculty, and students.

“This partnership fits with our strategy for the Boeing Aerospace and Autonomy Center at Kendall Square, which will advance the enabling technologies for autonomous aircraft,” says Greg Hyslop, Boeing chief technology officer and senior vice president of engineering, test, and technology. “We see great benefit in supporting MIT students challenged with developing real-world autonomy solutions.”

de MIT News http://bit.ly/2vkch8F

Top collegiate inventors awarded 2019 Lemelson-MIT Student Prize

Following a nationwide search for the most inventive college students, the Lemelson-MIT Program today announced the winners of the 2019 Lemelson-MIT Student Prize. The prize recognizes young inventors who have dedicated themselves to solving global problems. This year’s inventions range from innovative, low-cost cancer screening tools to an affordable clean water system, which ensures homes and families have clean, safe water. 

The Lemelson-MIT Student Prize is supported by The Lemelson Foundation, and serves as a catalyst for young inventors in the fields of health care, transportation and mobility, food/water and agriculture, and consumer devices. The program awarded a total of $90,000 in prizes to three undergraduate teams and four individual graduate student inventors, selected from a large and highly competitive pool of applicants from across the United States. Students were selected based on a variety of factors including: the overall inventiveness of their work, the invention’s potential for commercialization or adoption, and youth mentorship experience.

“We are inspired by the revolutionary work of this year’s winners. All of the inventions are designed with the intention of making the world a better place,” said Professor Michael J. Cima, faculty director of the Lemelson-MIT Program and associate dean of innovation within the MIT School of Engineering. “We are proud of how dedicated these young inventors are to combatting real-world problems.”

“We congratulate this year’s winners for their outstanding work tackling significant challenges in order to improve lives both in the United States and around the world,” said Carol Dahl, executive director at The Lemelson Foundation. “This diverse group of students drives home the opportunity that exists to inspire young minds across the country to create the essential inventions of today and tomorrow.”

2019 Lemelson-MIT Student Prize Winners 

The “Cure it!” Lemelson-MIT Student Prize rewards technology-based inventions that involve health care. The winners are:

• Mercy Asiedu of Duke University, $15,000 Graduate Winner


The majority of cervical cancer-related deaths occur in low and middle-income countries due to the lack of affordable screening technology. Mercy invented the Callascope, a high quality, low-cost, speculum-free device for cervical cancer screening and prevention. The device can be easily inserted into the vagina, like a tampon, either by a physician or for self-imaging/screening. It is fitted with a consumer-grade light source and camera to take images of the cervix from inside the body. The Callascope provides a cost-effective option for cervical cancer screenings in low-resource settings with limited available technologies. It can be connected to a mobile phone, tablet, or computer, and is coupled with an algorithm that uses machine learning to classify cervix images as normal or pre-cancerous.

• Ithemba: Laura Hinson, Madeline Lee, Sophia Triantis, and Valerie Zawicki of Johns Hopkins University, $10,000 Undergraduate Team Winner

The Ithemba team created a reusable, affordable, and contamination-free core needle breast biopsy device that is designed to support earlier breast cancer detection in low-resource settings. The reusable devices currently available on the market are expensive and require a 24-hour cleaning process. Ithemba’s novel device is not only affordable, but can also be sterilized instantly with a bleach wipe. With Ithemba’s device, performing breast biopsies will be significantly less expensive for hospitals and physicians in low-resource settings, and much safer for their patients.

The “Eat it!” Lemelson-MIT Student Prize rewards technology-based inventions that involve food/water and agriculture. The winners are:

• Julie Bliss Mullen of the University of Massachusetts at Amherst, $15,000 Graduate Winner



Mullen’s company, Aclarity LLC, offers a scalable electrochemical water purification technology marketed initially for residential use that uses low amounts of electricity to zap contaminants in water through advanced oxidation reactions. The technology disinfects pathogens, destroys organic contaminants, removes metals, and normalizes pH to produce truly clean and safe water. It reduces maintenance, uses low energy and purifies water faster and more efficiently than conventional treatment methods in the U.S. and globally.

• The BioEnergy Project: Enid Partika and William Tanaka of the University of California at San Diego, $10,000 Undergraduate Team Winner

The BioEnergy Project is a compact and scalable food-waste-to-food-and-fuel system that converts food waste from dining halls and restaurants into both nutrient-rich organic fertilizer that can be used to grow more food, as well as electricity that is generated from biogas. Right now, 40 percent of all food produced is wasted and dumped into landfills. When food decomposes in a landfill it generates methane, which is released into the atmosphere. Currently, food waste is responsible for 8 percent of the total anthropogenic greenhouse gas emissions globally. The BioEnergy Project’s invention is a cyclical system that can tackle the environmental and agricultural concerns of food insecurity, the need for renewable energy sources, and addresses climate change by capturing and utilizing a methane source that would otherwise be released into the atmosphere from landfills.

The “Move it!” Lemelson-MIT Student Prize rewards technology-based inventions that involve transportation and mobility. The winners are:

• Federico Scurti of North Carolina State University, $15,000 Graduate Winner

Scurti developed an internal monitoring system for high-temperature superconductors (HTS), consisting of a sensing system to detect local, incipient failures in the HTS wire that generates the magnetic field needed to operate electric motors or magnetic levitation (MagLev) trains. The sensing system is based on optical fibers embedded into superconducting wires that are able to prevent failure of the superconductor. This invention allows for reliable operation of HTS systems, thereby bringing HTS materials and systems to fruition via applications such as electric motors for carbon-free ships and aircrafts, carbon-free, high-speed MagLev trains, and nuclear fusion reactors for power generation.

• Portal Entryways: Josh Horne and Morgen Glessing of Brigham Young University, $10,000 Undergraduate Team Winner

Portal Entryways is a wireless device that opens disabled-accessible doors when a user approaches with the Portal smartphone application. A small wireless receiver is installed on the door and the user’s Portal app uses proximity to tell the door when to open upon approach. In addition to benefitting people with mobility-related disabilities, the system also enables facilities managers to track door usage data in order to maintain accessibility.

The “Use it!” Lemelson-MIT Student Prize rewards technology-based inventions that involve consumer devices. The winners are:

• Arnav Kapur of MIT, $15,000 Graduate Winner

The headset-like device, AlterEgo, is a sensory and auditory feedback system that uses neuromuscular signals from the brain’s speech system to extract speech. When we talk to ourselves internally, our brain transmits electrical signals to the vocal cords and internal muscles involved in speech production. With AlterEgo, an artificial intelligence agent is able to make sense of these signals and prepare a response. The user can hear the AI agent’s responses through vibrations in the skull and inner ear, thus making the process entirely internal. The AI agent can also send the information to a computer, to help an individual with a speech disability communicate in real-time.

Students interested in applying for the 2020 Lemelson-MIT Student Prize can find more information here. The 2020 Student Prize application will open in May.

de MIT News http://bit.ly/2DuwRaZ

Improving security as artificial intelligence moves to smartphones

Smartphones, security cameras, and speakers are just a few of the devices that will soon be running more artificial intelligence software to speed up image- and speech-processing tasks. A compression technique known as quantization is smoothing the way by making deep learning models smaller to reduce computation and energy costs. But smaller models, it turns out, make it easier for malicious attackers to trick an AI system into misbehaving — a concern as more complex decision-making is handed off to machines. 

In a new study, MIT and IBM researchers show just how vulnerable compressed AI models are to adversarial attack, and they offer a fix: add a mathematical constraint during the quantization process to reduce the odds that an AI will fall prey to a slightly modified image and misclassify what they see. 

When a deep learning model is reduced from the standard 32 bits to a lower bit length, it’s more likely to misclassify altered images due to an error amplification effect: The manipulated image becomes more distorted with each extra layer of processing. By the end, the model is more likely to mistake a bird for a cat, for example, or a frog for a deer.   

Models quantized to 8 bits or fewer are more susceptible to adversarial attacks, the researchers show, with accuracy falling from an already low 30-40 percent to less than 10 percent as bit width declines. But controlling the Lipschitz constraint during quantization restores some resilience. When the researchers added the constraint, they saw small performance gains in an attack, with the smaller models in some cases outperforming the 32-bit model. 

“Our technique limits error amplification and can even make compressed deep learning models more robust than full-precision models,” says Song Han, an assistant professor in MIT’s Department of Electrical Engineering and Computer Science and a member of MIT’s Microsystems Technology Laboratories. “With proper quantization, we can limit the error.” 

The team plans to further improve the technique by training it on larger datasets and applying it to a wider range of models. “Deep learning models need to be fast and secure as they move into a world of internet-connected devices,” says study coauthor Chuang Gan, a researcher at the MIT-IBM Watson AI Lab. “Our Defensive Quantization technique helps on both fronts.”

The researchers, who include MIT graduate student Ji Lin, present their results at the International Conference on Learning Representations in May.

In making AI models smaller so that they run faster and use less energy, Han is using AI itself to push the limits of model compression technology. In related recent work, Han and his colleagues show how reinforcement learning can be used to automatically find the smallest bit length for each layer in a quantized model based on how quickly the device running the model can process images. This flexible bit width approach reduces latency and energy use by as much as 200 percent compared to a fixed, 8-bit model, says Han. The researchers will present their results at the Computer Vision and Pattern Recognition conference in June. 

de MIT News http://bit.ly/2ISc8kS

Forums address MIT’s plans to reshape its computer science education

Community forums held last week for the MIT Stephen A. Schwarzman College of Computing working groups addressed how new ideas for its curriculum, research, infrastructure, and operations can best serve the community and society.

In February, MIT charged five working groups with generating ideas for the structure and operation of the new college. The groups are: Academic Degrees, Social Implications and Responsibilities of Computing, Organizational Structure, Faculty Appointments, and Computing Infrastructure. The groups each have two co-chairs and anywhere from about a dozen to more than 20 other members, representing many departments, labs, and centers across the Institute. The groups meet regularly, and will soon submit a report to MIT administration that outlines many potential ideas for their respective assignments.

To keep the community updated, three forums for all five working groups were held Wednesday, April 17, and Thursday, April 18. Co-chairs and members for each group delivered brief presentations on their group’s goals, information-gathering processes, current ideas, lessons learned, challenges, and questions left unanswered. Then, they opened the floor to questions and comments from the audience, which comprised dozens of students, faculty, and staff from across MIT.

In Wednesday morning’s forum, held in the Kresge Small Auditorium, Social Implications and Responsibilities of Computing co-chair Julie Shah shared a sentiment echoed by all working group presenters: The new college presents an opportunity to rethink how computer science is taught and researched at MIT and beyond.

“There is a sense we have a unique opportunity here in designing something across the educational curriculum, and also thinking about how we engage external parties — producers of technologies, consumers, and regulators,” said Shah, an associate professor of aeronautics and astronautics. “And what we do here can potentially be a model that lifts all of us that are working in the space across the country, and farther, together.”

All-inclusive computing

While the forums addressed a broad range of topics, common threads included: bringing all of MIT’s disciplines — including the humanities and social sciences — into the fold, lowering various barriers for computing, injecting ethics into research and coursework, and providing adequate computing resources for all.

In her presentation, Social Implications and Responsibilities co-chair Melissa Nobles said her group is considering new curricula for the 21st century that loop in ethical, social, and policy analyses. Their research shows current coursework isn’t necessarily optimal in that regard. She noted a need for investments of time and money to design new curricula.

“Certain classes do offer some discussion of ethics, but we think not nearly as robust … and certainly not as widespread as it will need to be,” said Nobles, who is the Kenan Sahin Dean and professor of political science in the School of Humanities, Arts, and Social Sciences. “We want students to think about their own particular role in [ethics] as creators of technologies … but also recognizing that we live within a larger economic, political, and social context. … I think the general idea is that it’s [about] individual behavior [and] the consequences of what we do.”

So far, a challenge for Academic Degrees, said co-chair Srini Devadas in the same forum, has been introducing computing to all MIT departments, especially the humanities and social sciences. The group, for instance, has looked into making MIT’s computer science minor more flexible for a greater diversity of students, and discussed possibilities for various computing graduate certificates. “We want to teach ground-breaking concepts, but I think it’s going to be a challenge [to serve] the entirety of the community of people interested in being in the computation core area or using it in their major discipline,” Devadas said. “This has to be as far away from a silo as we can possibly make it.”

In a similar vein, in Wednesday afternoon’s forum, also held in Kresge’s Small Auditorium, Organizational Structure co-chair Asu Ozdaglar said her group’s key goals right now are: ensuring the college nurtures interdisciplinary research, promoting computing across the MIT curriculum, and thoroughly integrating social sciences and humanities in with computing.

The group is also discussing the particularly challenging split affiliations in the Department of Electrical Engineering and Computer Science (EECS). “We have a spectrum of research, where it’s difficult to say where electrical engineering ends and computer science starts,” said Ozdaglar, the School of Engineering Distinguished Professor of Engineering and head of EECS. “We’re trying to figure out what design option [works best].”

In his presentation, Faculty Appointments co-chair Eran Ben-Joseph, head of the Department of Urban Studies and Planning, said one of the group’s guiding principles is attracting diverse faculty to create “an interdisciplinary hub to bring together unique faculty [and attracting] faculty that are different, and doing things that are nonconventional.”

One potential idea is hiring faculty in “clusters” across, say, science, engineering, and humanities that would teach and research a certain preselected topic that spans different disciplines. The idea is to “find a set of people who work collectively across boundaries and make them a group, and not just individuals working alone,” said group member Isaac Chuang, a professor of electrical engineering and a professor of physics.

Thursday morning’s forum, held on the sixth floor of the Samberg Center, saw Computing Infrastructure discuss best practices in hosting servers, storing and sharing tons of data, and cloud computing, among other factors. So far, the group has found broad support for creating a centralized network that provides “infinite” computing resources to everyone on campus. “[That] lowers the barrier of entry to places not necessarily involved in computing … and people that don’t particularly have access” to computing, said group co-chair Benoit Forget, an associate professor of nuclear science and engineering.

A major challenge, on the other hand, has been in setting up a system to ensure open-access data while navigating privacy, access, and licensing. In fact, a few audience members from different areas of MIT — including from the MIT Sloan School of Management, the Comparative Media Studies/Writing program, and the McGovern Institute for Brain Research at MIT — voiced concerns about the expenses and restrictions in accessing certain datasets for research. Currently, the working group is considering options for an Institute-wide data licensing and privacy framework that can address those issues.

Remaining open and dynamic

In addressing what will be in the working group reports, Shah drove home the point that the groups won’t propose specific recommendations, but instead will detail many different ideas to choose from and modify.

“Early in this process, it’s important for us to … think openly and understand the pros and cons of different approaches, so we can design a system that works well across all of MIT,” she said. Once groups submit reports, she added, there will be further discussions with the community: “This is not the end of input from you as a community — it’s the beginning.”

One notable question came from an audience member in Wednesday afternoon’s forum: After so much work, what if you get it wrong? The present co-chairs agreed that, while they hope to get many things right, there will always be room for improvement. Especially because the field of computer science so rapidly changes, the working groups must remain adaptable to fluctuations in academics and research, said Nelson Repenning, co-chair of Organizational Structure.

“Even if we do get it right now, the world’s going to change pretty quickly, and we are going to have to update [the structure and organization],” said Repenning, the associate dean of leadership and special projects and the Distinguished Professor of System Dynamics and Organization Studies at the MIT Sloan School of Management. “An element of dynamism to this is absolutely essential.”

de MIT News http://bit.ly/2UOrFt7