Artificial Intelligence

Whatever business a company may be in, software plays an increasingly vital role, from managing inventory to interfacing with customers. Software developers, as a result, are in greater demand than ever, and that’s driving the push to automate some of the easier tasks that take up their time. 

Productivity tools like Eclipse and Visual Studio suggest snippets of code that developers can easily drop into their work as they write. These automated features are powered by sophisticated language models that have learned to read and write computer code after absorbing thousands of examples. But like other deep learning models trained on big datasets without explicit instructions, language models designed for code-processing have baked-in vulnerabilities.

“Unless you’re really careful, a hacker can subtly manipulate inputs to these models to make them predict anything,” says Shashank Srikant, a graduate student in MIT’s Department of Electrical Engineering and Computer Science. “We’re trying to study and prevent that.”

In a new paper, Srikant and the MIT-IBM Watson AI Lab unveil an automated method for finding weaknesses in code-processing models, and retraining them to be more resilient against attacks. It’s part of a broader effort by MIT researcher Una-May O’Reilly and IBM-affiliated researcher Sijia Liu to harness AI to make automated programming tools smarter and more secure. The team will present its results next month at the International Conference on Learning Representations.

A machine capable of programming itself once seemed like science fiction. But an exponential rise in computing power, advances in natural language processing, and a glut of free code on the internet have made it possible to automate at least some aspects of software design. 

Trained on GitHub and other program-sharing websites, code-processing models learn to generate programs just as other language models learn to write news stories or poetry. This allows them to act as a smart assistant, predicting what software developers will do next, and offering an assist. They might suggest programs that fit the task at hand, or generate program summaries to document how the software works. Code-processing models can also be trained to find and fix bugs. But despite their potential to boost productivity and improve software quality, they pose security risks that researchers are just starting to uncover.

Srikant and his colleagues have found that code-processing models can be deceived simply by renaming a variable, inserting a bogus print statement, or introducing other cosmetic operations into programs the model tries to process. These subtly altered programs function normally, but dupe the model into processing them incorrectly, rendering the wrong decision.

The mistakes can have serious consequences for code-processing models of all types. A malware-detection model might be tricked into mistaking a malicious program for benign. A code-completion model might be duped into offering wrong or malicious suggestions. In both cases, viruses may sneak by the unsuspecting programmer. A similar problem plagues computer vision models: Edit a few key pixels in an input image and the model can confuse pigs for planes, and turtles for rifles, as other MIT research has shown. 

Like the best language models, code-processing models have one crucial flaw: They’re experts on the statistical relationships among words and phrases, but only vaguely grasp their true meaning. OpenAI’s GPT-3 language model, for example, can write prose that veers from eloquent to nonsensical, but only a human reader can tell the difference. 

Code-processing models are no different. “If they’re really learning intrinsic properties of the program, then it should be hard to fool them,” says Srikant. “But they’re not. They’re currently relatively easy to deceive.”

In the paper, the researchers propose a framework for automatically altering programs to expose weak points in the models processing them. It solves a two-part optimization problem; an algorithm identifies sites in a program where adding or replacing text causes the model to make the biggest errors. It also identifies what kinds of edits pose the greatest threat. 

What the framework reveals, the researchers say, is just how brittle some models are. Their text summarization model failed a third of the time when a single edit was made to a program; it failed more than half of the time when five edits were made, they report. On the flip side, they show that the model is able to learn from its mistakes, and in the process potentially gain a deeper understanding of programming.

“Our framework for attacking the model, and retraining it on those particular exploits, could potentially help code-processing models get a better grasp of the program’s intent,” says Liu, co-senior author of the study. “That’s an exciting direction waiting to be explored.”

In the background, a larger question remains: what exactly are these black-box deep-learning models learning? “Do they reason about code the way humans do, and if not, how can we make them?” says O’Reilly. “That’s the grand challenge ahead for us.” More

Over the past decade, hospitals and other health care providers have put massive amounts of time and energy into adopting electronic health care records, turning hastily scribbled doctors’ notes into durable sources of information. But collecting these data is less than half the battle. It can take even more time and effort to turn these records into actual insights — ones that use the learnings of the past to inform future decisions.

Cardea, a software system built by researchers and software engineers at MIT’s Data to AI Lab (DAI Lab), is built to help with that. By shepherding hospital data through an ever-increasing set of machine learning models, the system could assist hospitals in planning for events as large as global pandemics and as small as no-show appointments.

With Cardea, hospitals may eventually be able to solve “hundreds of different types of machine learning problems,” says Kalyan Veeramanchaneni, principal investigator of the DAI Lab and a principal research scientist in MIT’s Laboratory for Information and Decision Systems (LIDS). Because the framework is open-source, and uses generalizable techniques, they can also share these solutions with each other, increasing transparency and enabling teamwork.

Automated for the people

Cardea belongs to a field called automated machine learning, or AutoML. Machine learning is increasingly common, used for everything from drug development to credit card fraud detection. The goal of AutoML is to democratize these predictive tools, making it easier for people — including, eventually, non-experts — to build, use, and understand them, says Veeramachaneni.

Instead of requiring people to design and code an entire machine learning model, AutoML systems like Cardea surface existing ones, along with explanations of what they do and how they work. Users can then mix and match modules to accomplish their goals, like going to a buffet rather than cooking a meal from scratch.

For instance, data scientists have built a number of machine learning tools for health care, but most of them aren’t very accessible — even to experts. “They’re written up in papers and hidden away,” says Sarah Alnegheimish, a graduate student in LIDS. To build Cardea, she and her colleagues have been unearthing these tools and bringing them together, aiming to form “a powerful reference” for hospital problem-solvers, she says.

Step by step

To turn reams of data into useful predictions, Cardea walks users through a pipeline, with choices and safeguards at each step. They are first greeted by a data assembler, which ingests the information they provide. Cardea is built to work with Fast Healthcare Interoperability Resources (FHIR), the current industry standard for electronic health care records.

Hospitals vary in exactly how they use FHIR, so Cardea has been built to “adapt to different conditions and different datasets seamlessly,” says Veeramachaneni. If there are discrepancies within the data, Cardea’s data auditor points them out, so that they can be fixed or dismissed.

Next, Cardea asks the user what they want to find out. Perhaps they would like to estimate how long a patient might stay in the hospital. Even seemingly small questions like this one are crucial when it comes to day-to-day hospital operations — especially now, as health care facilities manage their resources during the Covid-19 pandemic, says Alnegheimish. Users can choose between different models, and the software system then uses the dataset and models to learn patterns from previous patients, and to predict what could happen in this case, helping stakeholders plan ahead.

Currently, Cardea is set up to help with four types of resource-allocation questions. But because the pipeline incorporates so many different models, it can be easily adapted to other scenarios that might arise. As Cardea grows, the goal is for stakeholders to eventually be able to use it to “solve any prediction problem within the health care domain,” Alnegheimish says.

The team presented their paper describing the system at the IEEE International Conference on Data Science and Advanced Analytics in October 2020. The researchers tested the accuracy of the system against users of a popular data science platform, and found that it out-competed 90 percent of them. They also tested its efficacy, asking data analysts to use Cardea to make predictions on a demo health care dataset. They found that Cardea significantly improved their efficiency — for example, feature engineering, which the analysts said usually takes them an average of two hours, took them five minutes instead.

Trust the process

Hospital workers are often tasked with making high-stakes, critical decisions. It’s vital that they trust any tools they use along the way, including Cardea. It’s not enough for users to plug in some numbers, press a button, and get an answer: “They should get some sense of the model, and they should know what is going on,” says Dongyu Liu, a postdoc in LIDS.

To build in even more transparency, Cardea’s next step is a model audit. Like all predictive apparatuses, machine learning models have strengths and weaknesses. By laying these out, Cardea gives the user the ability to decide whether to accept this model’s results, or to start again with a new one.

Cardea was released to the public earlier this year. Because it’s open source, users are welcome to integrate their own tools. The team also took pains to ensure that the software system is not only available, but understandable and easy to use. This will also help with reproducibility, Veeramachaneni says, so that predictions made on models built with the software can be understood and checked by others.  

The team also plans to build in more data visualizers and explanations, to provide an even deeper view, and make the software system more accessible to non-experts, Liu says.

“The hope is for people to adopt it, and start contributing to it,” Alnegheimish says. “With the help of the community, we can make it something much more powerful.”  More

Melanoma is a type of malignant tumor responsible for more than 70 percent of all skin cancer-related deaths worldwide. For years, physicians have relied on visual inspection to identify suspicious pigmented lesions (SPLs), which can be an indication of skin cancer. Such early-stage identification of SPLs in primary care settings can improve melanoma prognosis and significantly reduce treatment cost.

The challenge is that quickly finding and prioritizing SPLs is difficult, due to the high volume of pigmented lesions that often need to be evaluated for potential biopsies. Now, researchers from MIT and elsewhere have devised a new artificial intelligence pipeline, using deep convolutional neural networks (DCNNs) and applying them to analyzing SPLs through the use of wide-field photography common in most smartphones and personal cameras.

How it works: A wide-field image, acquired with a smartphone camera, shows large skin sections from a patient in a primary-care setting. An automated system detects, extracts, and analyzes all pigmented skin lesions observable in the wide-field image. A pre-trained deep convolutional neural network (DCNN) determines the suspiciousness of individual pigmented lesions and marks them (yellow = consider further inspection, red = requires further inspection or referral to dermatologist). Extracted features are used to further assess pigmented lesions and to display results in a heatmap format.

Animation courtesy of the researchers.

Previous item
Next item

DCNNs are neural networks that can be used to classify (or “name”) images to then cluster them (such as when performing a photo search). These machine learning algorithms belong to the subset of deep learning.

Using cameras to take wide-field photographs of large areas of patients’ bodies, the program uses DCNNs to quickly and effectively identify and screen for early-stage melanoma, according to Luis R. Soenksen, a postdoc and a medical device expert currently acting as MIT’s first Venture Builder in Artificial Intelligence and Healthcare. Soenksen conducted the research with MIT researchers, including MIT Institute for Medical Engineering and Science (IMES) faculty members Martha J. Gray, W. Kieckhefer Professor of Health Sciences and Technology, professor of electrical engineering and computer science; and James J. Collins, Termeer Professor of Medical Engineering and Science and Biological Engineering.

Soenksen, who is the first author of the recent paper, “Using Deep Learning for Dermatologist-level Detection of Suspicious Pigmented Skin Lesions from Wide-field Images,” published in Science Translational Medicine, explains that “Early detection of SPLs can save lives; however, the current capacity of medical systems to provide comprehensive skin screenings at scale are still lacking.”

The paper describes the development of an SPL analysis system using DCNNs to more quickly and efficiently identify skin lesions that require more investigation, screenings that can be done during routine primary care visits, or even by the patients themselves. The system utilized DCNNs to optimize the identification and classification of SPLs in wide-field images.

Using AI, the researchers trained the system using 20,388 wide-field images from 133 patients at the Hospital Gregorio Marañón in Madrid, as well as publicly available images. The images were taken with a variety of ordinary cameras that are readily available to consumers. Dermatologists working with the researchers visually classified the lesions in the images for comparison. They found that the system achieved more than 90.3 percent sensitivity in distinguishing SPLs from nonsuspicious lesions, skin, and complex backgrounds, by avoiding the need for cumbersome and time-consuming individual lesion imaging. Additionally, the paper presents a new method to extract intra-patient lesion saliency (ugly duckling criteria, or the comparison of the lesions on the skin of one individual that stand out from the rest) on the basis of DCNN features from detected lesions.

“Our research suggests that systems leveraging computer vision and deep neural networks, quantifying such common signs, can achieve comparable accuracy to expert dermatologists,” Soenksen explains. “We hope our research revitalizes the desire to deliver more efficient dermatological screenings in primary care settings to drive adequate referrals.”

Doing so would allow for more rapid and accurate assessments of SPLS and could lead to earlier treatment of melanoma, according to the researchers.

Gray, who is senior author of the paper, explains how this important project developed: “This work originated as a new project developed by fellows (five of the co-authors) in the MIT Catalyst program, a program designed to nucleate projects that solve pressing clinical needs. This work exemplifies the vision of HST/IMES devotee (in which tradition Catalyst was founded) of leveraging science to advance human health.” This work was supported by Abdul Latif Jameel Clinic for Machine Learning in Health and by the Consejería de Educación, Juventud y Deportes de la Comunidad de Madrid through the Madrid-MIT M+Visión Consortium. More

In recent years, robots have gained artificial vision, touch, and even smell. “Researchers have been giving robots human-like perception,” says MIT Associate Professor Fadel Adib. In a new paper, Adib’s team is pushing the technology a step further. “We’re trying to give robots superhuman perception,” he says.

The researchers have developed a robot that uses radio waves, which can pass through walls, to sense occluded objects. The robot, called RF-Grasp, combines this powerful sensing with more traditional computer vision to locate and grasp items that might otherwise be blocked from view. The advance could one day streamline e-commerce fulfillment in warehouses or help a machine pluck a screwdriver from a jumbled toolkit.

The research will be presented in May at the IEEE International Conference on Robotics and Automation. The paper’s lead author is Tara Boroushaki, a research assistant in the Signal Kinetics Group at the MIT Media Lab. Her MIT co-authors include Adib, who is the director of the Signal Kinetics Group; and Alberto Rodriguez, the Class of 1957 Associate Professor in the Department of Mechanical Engineering. Other co-authors include Junshan Leng, a research engineer at Harvard University, and Ian Clester, a PhD student at Georgia Tech.

Play video

As e-commerce continues to grow, warehouse work is still usually the domain of humans, not robots, despite sometimes-dangerous working conditions. That’s in part because robots struggle to locate and grasp objects in such a crowded environment. “Perception and picking are two roadblocks in the industry today,” says Rodriguez. Using optical vision alone, robots can’t perceive the presence of an item packed away in a box or hidden behind another object on the shelf — visible light waves, of course, don’t pass through walls.

But radio waves can.

For decades, radio frequency (RF) identification has been used to track everything from library books to pets. RF identification systems have two main components: a reader and a tag. The tag is a tiny computer chip that gets attached to — or, in the case of pets, implanted in — the item to be tracked. The reader then emits an RF signal, which gets modulated by the tag and reflected back to the reader.

The reflected signal provides information about the location and identity of the tagged item. The technology has gained popularity in retail supply chains — Japan aims to use RF tracking for nearly all retail purchases in a matter of years. The researchers realized this profusion of RF could be a boon for robots, giving them another mode of perception.

“RF is such a different sensing modality than vision,” says Rodriguez. “It would be a mistake not to explore what RF can do.”

RF Grasp uses both a camera and an RF reader to find and grab tagged objects, even when they’re fully blocked from the camera’s view. It consists of a robotic arm attached to a grasping hand. The camera sits on the robot’s wrist. The RF reader stands independent of the robot and relays tracking information to the robot’s control algorithm. So, the robot is constantly collecting both RF tracking data and a visual picture of its surroundings. Integrating these two data streams into the robot’s decision making was one of the biggest challenges the researchers faced.

“The robot has to decide, at each point in time, which of these streams is more important to think about,” says Boroushaki. “It’s not just eye-hand coordination, it’s RF-eye-hand coordination. So, the problem gets very complicated.”

The robot initiates the seek-and-pluck process by pinging the target object’s RF tag for a sense of its whereabouts. “It starts by using RF to focus the attention of vision,” says Adib. “Then you use vision to navigate fine maneuvers.” The sequence is akin to hearing a siren from behind, then turning to look and get a clearer picture of the siren’s source.

With its two complementary senses, RF Grasp zeroes in on the target object. As it gets closer and even starts manipulating the item, vision, which provides much finer detail than RF, dominates the robot’s decision making.

RF Grasp proved its efficiency in a battery of tests. Compared to a similar robot equipped with only a camera, RF Grasp was able to pinpoint and grab its target object with about half as much total movement. Plus, RF Grasp displayed the unique ability to “declutter” its environment — removing packing materials and other obstacles in its way in order to access the target. Rodriguez says this demonstrates RF Grasp’s “unfair advantage” over robots without penetrative RF sensing. “It has this guidance that other systems simply don’t have.”

RF Grasp could one day perform fulfilment in packed e-commerce warehouses. Its RF sensing could even instantly verify an item’s identity without the need to manipulate the item, expose its barcode, then scan it. “RF has the potential to improve some of those limitations in industry, especially in perception and localization,” says Rodriguez.

Adib also envisions potential home applications for the robot, like locating the right Allen wrench to assemble your Ikea chair. “Or you could imagine the robot finding lost items. It’s like a super-Roomba that goes and retrieves my keys, wherever the heck I put them.”

The research is sponsored by the National Science Foundation, NTT DATA, Toppan, Toppan Forms, and the Abdul Latif Jameel Water and Food Systems Lab (J-WAFS). More

In recent years there have been exciting breakthroughs in wearable technologies, like smartwatches that can monitor your breathing and blood oxygen levels. 

But what about a wearable that can detect how you move as you do a physical activity or play a sport, and could potentially even offer feedback on how to improve your technique? 

And, as a major bonus, what if the wearable were something you’d actually already be wearing, like a shirt of a pair of socks?

That’s the idea behind a new set of MIT-designed clothing that use special fibers to sense a person’s movement via touch. Among other things, the researchers showed that their clothes can actually determine things like if someone is sitting, walking, or doing particular poses.

The group from MIT’s Computer Science and Artificial Intelligence Lab (CSAIL) says that their clothes could be used for athletic training and rehabilitation. With patients’ permission, they could even help passively monitor the health of residents in assisted-care facilities and determine if, for example, someone has fallen or is unconscious.  

The researchers have developed a range of prototypes, from socks and gloves to a full vest. The team’s “tactile electronics” use a mix of more typical textile fibers alongside a small amount of custom-made functional fibers that sense pressure from the person wearing the garment.

According to CSAIL graduate student Yiyue Luo, a key advantage of the team’s design is that, unlike many existing wearable electronics, theirs can be incorporated into traditional large-scale clothing production. The machine-knitted tactile textiles are soft, stretchable, breathable, and can take a wide range of forms. 

“Traditionally it’s been hard to develop a mass-production wearable that provides high-accuracy data across a large number of sensors,” says Luo, lead author on a new paper about the project that is appearing in this month’s edition of Nature Electronics. “When you manufacture lots of sensor arrays, some of them will not work and some of them will work worse than others, so we developed a self-correcting mechanism that uses a self-supervised machine learning algorithm to recognize and adjust when certain sensors in the design are off-base.”

The team’s clothes have a range of capabilities. Their socks predict motion by looking at how different sequences of tactile footprints correlate to different poses as the user transitions from one pose to another. The full-sized vest can also detect the wearers’ pose, activity, and the texture of the contacted surfaces.

The authors imagine a coach using the sensor to analyze people’s postures and give suggestions on improvement. It could also be used by an experienced athlete to record their posture so that beginners can learn from them. In the long term, they even imagine that robots could be trained to learn how to do different activities using data from the wearables. 

“Imagine robots that are no longer tactilely blind, and that have ‘skins’ that can provide tactile sensing just like we have as humans,” says corresponding author Wan Shou, a postdoc at CSAIL. “Clothing with high-resolution tactile sensing opens up a lot of exciting new application areas for researchers to explore in the years to come.”

The paper was co-written by MIT professors Antonio Torralba, Wojciech Matusik, and Tomás Palacios, alongside PhD students Yunzhu Li, Pratyusha Sharma, and Beichen Li; postdoc Kui Wu; and research engineer Michael Foshey. 

The work was partially funded by Toyota Research Institute. More

Small-molecule therapeutics treat a wide variety of diseases, but their effectiveness is often diminished because of their pharmacokinetics — what the body does to a drug. After administration, the body dictates how much of the drug is absorbed, which organs the drug enters, and how quickly the body metabolizes and excretes the drug again.

Nanoparticles, usually made out of lipids, polymers, or both, can improve the pharmacokinetics, but they can be complex to produce and often carry very little of the drug.

Some combinations of small-molecule cancer drugs and two small-molecule dyes have been shown to self-assemble into nanoparticles with extremely high payloads of drugs, but it is difficult to predict which small-molecule partners will form nanoparticles among the millions of possible pairings.

MIT researchers have developed a screening platform that combines machine learning with high-throughput experimentation to identify self-assembling nanoparticles quickly. In a study published in Nature Nanotechnology, researchers screened 2.1 million pairings of small-molecule drugs and “inactive” drug ingredients, identifying 100 new nanoparticles with potential applications that include the treatment of cancer, asthma, malaria, and viral and fungal infections.

“We have previously described some of the negative and positive effects that inactive ingredients can have on drugs, and here, through a concerted collaboration across our laboratories and core facilities, describe an approach focusing on the potential positive effects these can have on nanoformulation,” says Giovanni Traverso, the Karl Van Tassel (1925) Career Development Professor of Mechanical Engineering, and senior corresponding author of the study.

Their findings point to a strategy for that solves for both the complexity of producing nanoparticles and the difficulty of loading large amounts of drugs onto them.

“So many drugs out there don’t live up to their full potential because of insufficient targeting, low bioavailability, or rapid drug metabolism,” says Daniel Reker, lead author of the study and a former postdoc in the laboratory of Robert Langer. “By working at the interface of data science, machine learning, and drug delivery, our hope is to rapidly expand our tool set for making sure a drug gets to the place it needs to be and can actually treat and help a human being.”

Langer, the David H. Koch Institute Professor at MIT and a member of the Koch Institute for Integrative Cancer Research, is also a senior author of the paper.

A cancer therapy meets its match

In order to develop a machine learning algorithm capable of identifying self-assembling nanoparticles, researchers first needed to build a dataset on which the algorithm could train. They selected 16 self-aggregating small-molecule drugs with a variety of chemical structures and therapeutic applications and a diverse set of 90 widely available compounds, including ingredients that are already added to drugs to make them taste better, last longer, or make them more stable. Because both the drugs and the inactive ingredients are already FDA-approved, the resulting nanoparticles are likely to be safer and move through the FDA approval process more quickly.

The team then tested every combination of small-molecule drug and inactive ingredient, enabled by the Swanson Biotechnology Center, a suite of core facilities providing advanced technical services within the Koch Institute. After mixing pairings and loading 384 samples at a time onto nanowell plates using robotics in the High Throughput Sciences core, researchers walked the plates, often with quickly degrading samples, next door to the Peterson (1957) Nanotechnology Materials Core Facility core to measure the size of particles with high throughput dynamic light scattering.

Now trained on 1,440 data points (with 94 nanoparticles already identified), the machine learning platform could be turned on a much bigger library of compounds. Screening 788 small-molecule drugs against more than 2,600 inactive drug ingredients, the platform identified 38,464 potential self-assembling nanoparticles from 2.1 million possible combinations.

The researchers selected six nanoparticles for further validation, including one composed of sorafenib, a treatment commonly used for advanced liver and other cancers, and glycyrrhizin, a compound frequently used as both a food and drug additive and most commonly known as licorice flavoring. Although sorafenib is the standard of care for advanced liver cancer, its effectiveness is limited.

In human liver cancer cell cultures, the sorafenib-glycyrrhizin nanoparticles worked twice as well as sorafenib by itself because more of the drug could enter the cells. Working with the Preclinical Modeling, Imaging and Testing facility at the Koch Institute, researchers treated mouse models of liver cancer to compare the effects of sorafenib-glycyrrhizin nanoparticles versus either compound by itself. They found that the nanoparticle significantly reduced levels of a marker associated with liver cancer progression compared to mice given sorafenib alone, and lived longer than mice given sorafenib or glycyrrhizin alone. The sorafenib-glycyrrhizin nanoparticle also showed improved targeting to the liver when compared to oral delivery of sorafenib, the current standard in the clinic, or when injecting sorafenib after it has been combined with cremophor, a commonly-used drug vehicle that improves water solubility but has toxic side effects.

Personalized drug delivery

The new platform may have useful applications beyond optimizing the efficiency of active drugs: it could be used to customize inactive compounds to suit the needs of individual patients. In earlier work, members of the team found that inactive ingredients could provoke adverse allergic reactions in some patients. Now, with the expanded machine learning toolbox, more options could be generated to provide alternatives for these patients.

“We have an opportunity to think about matching the delivery system to the patient,” explains Reker, now an assistant professor of biomedical engineering at Duke University. “We can account for things like drug absorption, genetics, even allergies to reduce side effects upon delivery. Whatever the mutation or medical condition, the right drug is only the right drug if it actually works for the patient.”

The tools for safe, efficacious drug delivery exist, but putting all the ingredients together can be a slow process. The combination of machine learning, rapid screening, and the ability to predict interactions among different combinations of materials will accelerate the design of drugs and the nanoparticles used to deliver them throughout the body.

In ongoing work, the team is looking not just to improve effective delivery of drugs but also for opportunities to create medications for people for whom standard formulations are not a good option, using big data to solve problems in small populations by looking at genetic history, allergies, and food reactions. More

Materials called perovskites are widely heralded as a likely replacement for silicon as the material of choice for solar cells, but their greatest drawback is their tendency to degrade relatively rapidly. Over recent years, the usable lifetime of perovskite-based cells has gradually improved from minutes to months, but it still lags far behind the decades expected from silicon, the material currently used for virtually all commercial solar panels.

Now, an international interdisciplinary team led by MIT has come up with a new approach to narrowing the search for the best candidates for long-lasting perovskite formulations, out of a vast number of potential combinations. Already, their system has zeroed in on one composition that in the lab has improved on existing versions more than tenfold. Even under real-world conditions at full solar cell level, beyond just a small sample in a lab, this type of perovskite has performed three times better than the state-of-the-art formulations.

The findings appear in the journal Matter, in a paper by MIT research scientist Shijing Sun, MIT professors, Moungi Bawendi,  John Fisher, and Tonio Buonassisi, who is also a principal investigator at the Singapore-MIT Alliance for Research and Technology (SMART), and 16 others from MIT, Germany, Singapore, Colorado, and New York.

Perovskites are a broad class of materials characterized by the way atoms are arranged in their layered crystal lattice. These layers, described by convention as A, B, and X, can each consist of a variety of different atoms or compounds. So, searching through the entire universe of such combinations to find the best candidates to meet specific goals — longevity, efficiency, manufacturability, and availability of source materials — is a slow and painstaking process, and largely one without any map for guidance.

“If you consider even just three elements, the most common ones in perovskites that people sub in and out are on the A site of the perovskite crystal structure,” which can each easily be varied by 1-percent increments in their relative composition, Buonassisi says. “The number of steps becomes just preposterous. It becomes very, very large” and thus impractical to search through systematically. Each step involves the complex synthesis process of creating a new material and then testing its degradation, which even under accelerated aging conditions is a time-consuming process.

The key to the team’s success is what they describe as a data fusion approach. This iterative method uses an automated system to guide the production and testing of a variety of formulations, then uses machine learning to go through the results of those tests, combined again with first-principles physical modeling, to guide the next round of experiments. The system keeps repeating that process, refining the results each time.

Buonassisi likes to compare the vast realm of possible compositions to an ocean, and he says most researchers have stayed very close to the shores of known formulations that have achieved high efficiencies, for example, by tinkering just slightly with those atomic configurations. However, “once in a while, somebody makes a mistake or has a stroke of genius and departs from that and lands somewhere else in composition space, and hey, it works better! A happy bit of serendipity, and then everybody moves over there” in their research. “But it’s not usually a structured thought process.”

This new approach, he says, provides a way to explore far offshore areas in search of better properties, in a more systematic and efficient way. In their work so far, by synthesizing and testing less than 2 percent of the possible combinations among three components, the researchers were able to zero in on what seems to be the most durable formulation of a perovskite solar cell material found to date.

“This story is really about the fusion of all the different sets of tools” used to find the new formulation, says Sun, who coordinated the international team that carried out the work, including the development of a high-throughput automated degradation test system that monitors the breakdown of the material through its changes in color as it darkens. To confirm the results, the team went beyond making a tiny chip in the lab and incorporated the material into a working solar cell.

“Another point of this work is that we actually demonstrate, all the way from the chemical selection until we actually make a solar cell in the end,” she says. “And it tells us that the machine-learning-suggested chemical is not only stable in its own freestanding form. They can also be translated into real-life solar cells, and they lead to improved reliability.” Some of their lab-scale demonstrations achieved longevity as much as 17 times greater than the baseline formula they started with, but even the full-cell demonstration, which includes the necessary interconnections, outlasted the existing materials by more than three times, she says.

Buonassisi says the method the team developed could also be applied to other areas of materials research involving similarly large ranges of choice in composition. “It really opens the door for a mode of research where you can have these short, quick loops of innovation happening, maybe at a subcomponent or a material level. And then once you zero in on the right composition, you bump it up into a longer loop that involves device fabrication, and you test it out” at that next level.

“It’s one of the big promises of the field to be able to do this type of work,” he says. “To see it actually happen was one of those [highly memorable] moments. I remember the exact place I was when I received the call from Shijing about these results — when you start to actually see these ideas come to life. It was really stunning.”

“What is particularly exciting about [this] advance is that the authors use physics to guide the intuition of the [optimization] process, rather than limiting the search space with hard constraints,” says University Professor Edward Sargent of the University of Toronto, a specialist in nanotechnology who was not connected with this research. “This approach will see widespread exploitation as machine learning continues to move toward solving real problems in materials science.”

The team included researchers at MIT, the Helmholz Intitute in Germany, the Colorado School of Mines, Brookhaven National Laboratory in New York, the Singapore-MIT Alliance for Research and Technology, and the Institute of Materials for Electronics and Energy Technology in Erlangen, Germany. The work was supported by DARPA, Total SA, the National Science Foundation, and the Skoltech NGP program. More

The potential of artificial intelligence to bring equity in health care has spurred significant research efforts. Racial, gender, and socioeconomic disparities have traditionally afflicted health care systems in ways that are difficult to detect and quantify. New AI technologies, however, are providing a platform for change.

Regina Barzilay, the School of Engineering Distinguished Professor of AI and Health and faculty co-lead of AI for the MIT Jameel Clinic; Fotini Christia, professor of political science and director of the MIT Sociotechnical Systems Research Center; and Collin Stultz, professor of electrical engineering and computer science and a cardiologist at Massachusetts General Hospital — discuss here the role of AI in equitable health care, current solutions, and policy implications. The three are co-chairs of the AI for Healthcare Equity Conference, taking place April 12.

Q: How can AI help address racial, gender, and socioeconomic disparities in health-care systems?

Stultz: Many factors contribute to economic disparities in health care systems. For one, there is little doubt that inherent human bias contributes to disparate health outcomes in marginalized populations. Although bias is an inescapable part of the human psyche, it is insidious, pervasive, and hard to detect. Individuals, in fact, are notoriously poor at detecting preexisting bias in their own perception of the world — a fact that has driven the development of implicit association tests that allow one to understand how underlying bias can affect decision-making.  

AI provides a platform for the development of methods that can make personalized medicine a reality, thereby ensuring that clinical decisions are made objectively with the goal of minimizing adverse outcomes across different populations. Machine learning, in particular, describes a set of methods that help computers learn from data. In principle, these methods can offer unbiased predictions that are based only on objective analyses of the underlying data.

Unfortunately, however, bias not only affects how individuals perceive the world around them, it also influences the datasets we use to build models. Observational datasets that store patient features and outcomes often reflect the underlying bias of health care providers; e.g., certain treatments may be preferentially offered to those who have high socioeconomic status. In short, algorithms can inherit our own biases. Making personalized medicine a reality is therefore predicated on our ability to develop and deploy unbiased tools that learn the patient-specific decisions from observational clinical data. Central to the success of this endeavor is the development of methods that can identify algorithmic bias and suggest mitigation strategies when bias is identified.

Informed, objective, and patient-specific clinical decisions are the future of modern clinical care. Machine learning will go a long way to making this a reality — achieving data-driven clinical insights devoid of implicit prejudice that can influence health-care decisions.

Q: What are some current AI solutions being developed in this space?

Barzilay: In most cases, biased predictions can be attributed to distributional properties of the training data. For instance, when some population is underrepresented in the training data, the resulting classifier is likely to underperform on this group. By default, models are optimized for the overall performance, thus inadvertently preferring to fit the majority class, at the expense of the rest. If we are aware of such minority groups in the data, we have multiple means to steer our learning algorithm towards fair behavior. For example, we can modify the learning objective where we enforce consistent accuracy across different groups, or reweigh the significance of training examples, amplifying “the voice” of the minority group.

Another common source of bias relates to “nuisance variations” where classification labels exhibit idiosyncratic correlations with some input features which are dataset-specific and are unlikely to generalize. In one infamous dataset with such property, health status of patients with the same medical history depended on their race. This bias was an unfortunate artifact of the way training data was constructed, but it resulted in systematic discrimination of Black patients. If such biases are known beforehand, we can mitigate their effect by forcing the model to reduce the effect of such attributes. In many cases though, biases of our training data are unknown. It is safe to assume that the environment in which the model will be applied is likely to exhibit some distributional divergence from the training data. To improve a model’s tolerance to such shifts, a number of approaches (like invariant risk minimization) explicitly train the model to robustly generalize to new environments.

However, we should be aware that algorithms are not magic wands that can correct all wrongs in messy, real-world training data. This is especially true when we are not aware of the peculiarity of a specific dataset. The latter scenario is unfortunately common in the health care domain where data curation and machine learning are often performed by different teams. These “hidden” biases have already resulted in deployed AI tools that systematically err on certain populations (like the model described above). In such cases, it is essential to provide physicians with tools that enable them to understand the rationale behind model predictions and detect biased predictions as soon as possible. A large body of work in machine learning is dedicated today to developing transparent models that can communicate their internal reasoning to users. At this point, our understanding of what types of rationales are particularly useful for doctors is limited, since AI tools are not yet part of routine medical practice. Therefore, one of the key goals of MIT’s Jameel Clinic is to deploy clinical AI algorithms in hospitals around the world and empirically study their performance in different populations and clinical settings. This data will inform the development of the next generation of self-explainable and fair AI tools.  

Q: What are the policy implications for government agencies and the industry of more equitable AI for health care?

Christia: The use of AI in health care is now a reality and for government agencies and the industry to reap the benefits of a more equitable AI for health care, they need to create an AI ecosystem. They have to work together closely and engage with clinicians and patients to prioritize the quality of the AI tools that get employed in this space, making sure they are safe and ready for prime-time. This means that AI tools that get deployed have to be well-tested and to lead to improvements in both clinician capacity and patient experience.

To that effect, government and industry players need to think about educational campaigns that inform health practitioners of the importance of specific AI interventions in complementing and augmenting their work to address equity. Beyond clinicians, there also has to be a focus on building confidence with minority patients that the introduction of these AI tools will result in overall better and more equitable care. It is particularly important to also be transparent about what the use of AI in health means for the individual patient, as well as assuage data privacy concerns of patients from minority populations who often lack trust in a “well-intentioned” health care system, given historical transgressions against them.

In the regulatory realm, government agencies would need to put together a framework that would allow them to have clarity over AI funding and liability with the industry and health care professionals so the highest-quality AI tools get deployed while also minimizing the associated risks for clinicians and patients using them. Regulations would need to make clear that the clinicians are not fully outsourcing their responsibility to the machine and outline the levels of professional accountability for their patients’ health. Working closely with the industry, clinicians and patients, government agencies would also have to monitor through data and patient experience the actual effectiveness of AI tools in addressing health care disparities on the ground, and be attuned to improving them. More