More stories

  • in

    3 Questions: A new PhD program from the Center for Computational Science and Engineering

    This fall, the Center for Computational Science and Engineering (CCSE), an academic unit in the MIT Schwarzman College of Computing, is introducing a new standalone PhD degree program that will enable students to pursue research in cross-cutting methodological aspects of computational science and engineering. The launch follows approval of the center’s degree program proposal at the May 2023 Institute faculty meeting.

    Doctoral-level graduate study in computational science and engineering (CSE) at MIT has, for the past decade, been offered through an interdisciplinary program in which CSE students are admitted to one of eight participating academic departments in the School of Engineering or School of Science. While this model adds a strong disciplinary component to students’ education, the rapid growth of the CSE field and the establishment of the MIT Schwarzman College of Computing have prompted an exciting expansion of MIT’s graduate-level offerings in computation.

    The new degree, offered by the college, will run alongside MIT’s existing interdisciplinary offerings in CSE, complementing these doctoral training programs and preparing students to contribute to the leading edge of the field. Here, CCSE co-directors Youssef Marzouk and Nicolas Hadjiconstantinou discuss the standalone program and how they expect it to elevate the visibility and impact of CSE research and education at MIT.

    Q: What is computational science and engineering?

    Marzouk: Computational science and engineering focuses on the development and analysis of state-of-the-art methods for computation and their innovative application to problems of science and engineering interest. It has intellectual foundations in applied mathematics, statistics, and computer science, and touches the full range of science and engineering disciplines. Yet, it synthesizes these foundations into a discipline of its own — one that links the digital and physical worlds. It’s an exciting and evolving multidisciplinary field.

    Hadjiconstantinou: Examples of CSE research happening at MIT include modeling and simulation techniques, the underlying computational mathematics, and data-driven modeling of physical systems. Computational statistics and scientific machine learning have become prominent threads within CSE, joining high-performance computing, mathematically-oriented programming languages, and their broader links to algorithms and software. Application domains include energy, environment and climate, materials, health, transportation, autonomy, and aerospace, among others. Some of our researchers focus on general and widely applicable methodology, while others choose to focus on methods and algorithms motivated by a specific domain of application.

    Q: What was the motivation behind creating a standalone PhD program?

    Marzouk: The new degree focuses on a particular class of students whose background and interests are primarily in CSE methodology, in a manner that cuts across the disciplinary research structure represented by our current “with-departments” degree program. There is a strong research demand for such methodologically-focused students among CCSE faculty and MIT faculty in general. Our objective is to create a targeted, coherent degree program in this field that, alongside our other thriving CSE offerings, will create the leading environment for top CSE students worldwide.

    Hadjiconstantinou: One of CCSE’s most important functions is to recruit exceptional students who are trained in and want to work in computational science and engineering. Experience with our CSE master’s program suggests that students with a strong background and interests in the discipline prefer to apply to a pure CSE program for their graduate studies. The standalone degree aims to bring these students to MIT and make them available to faculty across the Institute.

    Q: How will this impact computing education and research at MIT? 

    Hadjiconstantinou: We believe that offering a standalone PhD program in CSE alongside the existing “with-departments” programs will significantly strengthen MIT’s graduate programs in computing. In particular, it will strengthen the methodological core of CSE research and education at MIT, while continuing to support the disciplinary-flavored CSE work taking place in our participating departments, which include Aeronautics and Astronautics; Chemical Engineering; Civil and Environmental Engineering; Materials Science and Engineering; Mechanical Engineering; Nuclear Science and Engineering; Earth, Atmospheric and Planetary Sciences; and Mathematics. Together, these programs will create a stronger CSE student cohort and facilitate deeper exchanges between the college and other units at MIT.

    Marzouk: In a broader sense, the new program is designed to help realize one of the key opportunities presented by the college, which is to create a richer variety of graduate degrees in computation and to involve as many faculty and units in these educational endeavors as possible. The standalone CSE PhD will join other distinguished doctoral programs of the college — such as the Department of Electrical Engineering and Computer Science PhD; the Operations Research Center PhD; and the Interdisciplinary Doctoral Program in Statistics and the Social and Engineering Systems PhD within the Institute for Data, Systems, and Society — and grow in a way that is informed by them. The confluence of these academic programs, and natural synergies among them, will make MIT quite unique. More

  • in

    On the hunt for sustainable materials

    By the time she started high school, Avni Singhal had attended six different schools in a variety of settings, from a traditional public school to a self-paced program. The transitions opened her eyes to how widely educational environments can vary, and made her think about that impact on students.

    “Experiencing so many different types of educational systems exposed me to different ways of looking at things and how that shapes people’s worldviews,” says Singhal.

    Now a fourth-year PhD student in the Department of Materials Science and Engineering, Singhal is still thinking about increasing opportunities for her fellow students, while also pursuing her research. She devotes herself to both developing sustainable materials and improving the graduate experience in her department.

    She recently completed her two-year term as a student representative on the department’s graduate studies committee. In this role, she helped revamp the communication around the qualifying exams and introducing student input to the faculty search process.

    “It’s given me a lot of insight into how our department works,” says Singhal. “It’s a chance to get to know faculty, bring up issues that students experience, and work on changing things that we think could be improved.”

    At the same time, Singhal uses atomistic simulations to model material properties, with an eye toward sustainability. She is a part of the Learning Matter Lab, a group that merges data science tools with engineering and physics-based simulation to better design and understand materials. As part of a computational group, Singhal has worked on a range of projects in collaboration with other labs that are looking to combine computing with other disciplines. Some of this work is sponsored by the MIT Climate and Sustainability Consortium, which facilitates connections across MIT labs and industry.

    Joining the Learning Matter Lab was a step out of Singhal’s comfort zone. She arrived at MIT from the University of California at Berkeley with a joint degree in materials science and bioengineering, as well as a degree in electrical engineering and computer science.

    “I was generally interested in doing work on environment-related applications,” says Singhal. “I was pretty hesitant at first to switch entirely to computation because it’s a very different type of lifestyle of research than what I was doing before.”

    Singhal has taken the challenge in stride, contributing to projects including improving carbon capture molecules and developing new deconstructable, degradable plastics. Not only does Singhal have to understand the technical details of her own work, she also needs to understand the big picture and how to best wield the expertise of her collaborators.

    “When I came in, I was very wide-eyed, thinking computation can do everything because I had never done it before,” says Singhal. “It’s that curve where you know a little bit about something, and you think it can do everything. And then as you learn more, you learn where it can and can’t help us, where it can be valuable, and how to figure out in what part of a project it’s useful.”

    Singhal applies a similarly critical lens when thinking about graduate school as a whole. She notes that access to information and resources is often the main factor determining who enters selective educational programs, and that such access becomes increasingly limited at the graduate level.

    “I realized just how much applying is a function of knowing how to do it,” says Singhal, who co-organized and volunteers with the DMSE Application Assistance Program. The program matches prospective applicants with current students to give feedback on their application materials and provide insight into what it’s like attending MIT. Some of the first students Singhal mentored through the program are now also participants as well.

    “The further you get in your educational career, the more you realize how much assistance you got along the way to get where you are,” says Singhal. “That happens at every stage.”

    Looking toward the future, Singhal wants to continue to pursue research with a sustainability impact. She also wants to continue mentoring in some capacity but isn’t in a rush to figure out exactly what that will look like.

    “Grad school doesn’t mean I have to do one thing. I can stay open to all the possibilities of what comes next.”  More

  • in

    How an archeological approach can help leverage biased data in AI to improve medicine

    The classic computer science adage “garbage in, garbage out” lacks nuance when it comes to understanding biased medical data, argue computer science and bioethics professors from MIT, Johns Hopkins University, and the Alan Turing Institute in a new opinion piece published in a recent edition of the New England Journal of Medicine (NEJM). The rising popularity of artificial intelligence has brought increased scrutiny to the matter of biased AI models resulting in algorithmic discrimination, which the White House Office of Science and Technology identified as a key issue in their recent Blueprint for an AI Bill of Rights. 

    When encountering biased data, particularly for AI models used in medical settings, the typical response is to either collect more data from underrepresented groups or generate synthetic data making up for missing parts to ensure that the model performs equally well across an array of patient populations. But the authors argue that this technical approach should be augmented with a sociotechnical perspective that takes both historical and current social factors into account. By doing so, researchers can be more effective in addressing bias in public health. 

    “The three of us had been discussing the ways in which we often treat issues with data from a machine learning perspective as irritations that need to be managed with a technical solution,” recalls co-author Marzyeh Ghassemi, an assistant professor in electrical engineering and computer science and an affiliate of the Abdul Latif Jameel Clinic for Machine Learning in Health (Jameel Clinic), the Computer Science and Artificial Intelligence Laboratory (CSAIL), and Institute of Medical Engineering and Science (IMES). “We had used analogies of data as an artifact that gives a partial view of past practices, or a cracked mirror holding up a reflection. In both cases the information is perhaps not entirely accurate or favorable: Maybe we think that we behave in certain ways as a society — but when you actually look at the data, it tells a different story. We might not like what that story is, but once you unearth an understanding of the past you can move forward and take steps to address poor practices.” 

    Data as artifact 

    In the paper, titled “Considering Biased Data as Informative Artifacts in AI-Assisted Health Care,” Ghassemi, Kadija Ferryman, and Maxine Mackintosh make the case for viewing biased clinical data as “artifacts” in the same way anthropologists or archeologists would view physical objects: pieces of civilization-revealing practices, belief systems, and cultural values — in the case of the paper, specifically those that have led to existing inequities in the health care system. 

    For example, a 2019 study showed that an algorithm widely considered to be an industry standard used health-care expenditures as an indicator of need, leading to the erroneous conclusion that sicker Black patients require the same level of care as healthier white patients. What researchers found was algorithmic discrimination failing to account for unequal access to care.  

    In this instance, rather than viewing biased datasets or lack of data as problems that only require disposal or fixing, Ghassemi and her colleagues recommend the “artifacts” approach as a way to raise awareness around social and historical elements influencing how data are collected and alternative approaches to clinical AI development. 

    “If the goal of your model is deployment in a clinical setting, you should engage a bioethicist or a clinician with appropriate training reasonably early on in problem formulation,” says Ghassemi. “As computer scientists, we often don’t have a complete picture of the different social and historical factors that have gone into creating data that we’ll be using. We need expertise in discerning when models generalized from existing data may not work well for specific subgroups.” 

    When more data can actually harm performance 

    The authors acknowledge that one of the more challenging aspects of implementing an artifact-based approach is being able to assess whether data have been racially corrected: i.e., using white, male bodies as the conventional standard that other bodies are measured against. The opinion piece cites an example from the Chronic Kidney Disease Collaboration in 2021, which developed a new equation to measure kidney function because the old equation had previously been “corrected” under the blanket assumption that Black people have higher muscle mass. Ghassemi says that researchers should be prepared to investigate race-based correction as part of the research process. 

    In another recent paper accepted to this year’s International Conference on Machine Learning co-authored by Ghassemi’s PhD student Vinith Suriyakumar and University of California at San Diego Assistant Professor Berk Ustun, the researchers found that assuming the inclusion of personalized attributes like self-reported race improve the performance of ML models can actually lead to worse risk scores, models, and metrics for minority and minoritized populations.  

    “There’s no single right solution for whether or not to include self-reported race in a clinical risk score. Self-reported race is a social construct that is both a proxy for other information, and deeply proxied itself in other medical data. The solution needs to fit the evidence,” explains Ghassemi. 

    How to move forward 

    This is not to say that biased datasets should be enshrined, or biased algorithms don’t require fixing — quality training data is still key to developing safe, high-performance clinical AI models, and the NEJM piece highlights the role of the National Institutes of Health (NIH) in driving ethical practices.  

    “Generating high-quality, ethically sourced datasets is crucial for enabling the use of next-generation AI technologies that transform how we do research,” NIH acting director Lawrence Tabak stated in a press release when the NIH announced its $130 million Bridge2AI Program last year. Ghassemi agrees, pointing out that the NIH has “prioritized data collection in ethical ways that cover information we have not previously emphasized the value of in human health — such as environmental factors and social determinants. I’m very excited about their prioritization of, and strong investments towards, achieving meaningful health outcomes.” 

    Elaine Nsoesie, an associate professor at the Boston University of Public Health, believes there are many potential benefits to treating biased datasets as artifacts rather than garbage, starting with the focus on context. “Biases present in a dataset collected for lung cancer patients in a hospital in Uganda might be different from a dataset collected in the U.S. for the same patient population,” she explains. “In considering local context, we can train algorithms to better serve specific populations.” Nsoesie says that understanding the historical and contemporary factors shaping a dataset can make it easier to identify discriminatory practices that might be coded in algorithms or systems in ways that are not immediately obvious. She also notes that an artifact-based approach could lead to the development of new policies and structures ensuring that the root causes of bias in a particular dataset are eliminated. 

    “People often tell me that they are very afraid of AI, especially in health. They’ll say, ‘I’m really scared of an AI misdiagnosing me,’ or ‘I’m concerned it will treat me poorly,’” Ghassemi says. “I tell them, you shouldn’t be scared of some hypothetical AI in health tomorrow, you should be scared of what health is right now. If we take a narrow technical view of the data we extract from systems, we could naively replicate poor practices. That’s not the only option — realizing there is a problem is our first step towards a larger opportunity.”  More

  • in

    How machine learning models can amplify inequities in medical diagnosis and treatment

    Prior to receiving a PhD in computer science from MIT in 2017, Marzyeh Ghassemi had already begun to wonder whether the use of AI techniques might enhance the biases that already existed in health care. She was one of the early researchers to take up this issue, and she’s been exploring it ever since. In a new paper, Ghassemi, now an assistant professor in MIT’s Department of Electrical Science and Engineering (EECS), and three collaborators based at the Computer Science and Artificial Intelligence Laboratory, have probed the roots of the disparities that can arise in machine learning, often causing models that perform well overall to falter when it comes to subgroups for which relatively few data have been collected and utilized in the training process. The paper — written by two MIT PhD students, Yuzhe Yang and Haoran Zhang, EECS computer scientist Dina Katabi (the Thuan and Nicole Pham Professor), and Ghassemi — was presented last month at the 40th International Conference on Machine Learning in Honolulu, Hawaii.

    In their analysis, the researchers focused on “subpopulation shifts” — differences in the way machine learning models perform for one subgroup as compared to another. “We want the models to be fair and work equally well for all groups, but instead we consistently observe the presence of shifts among different groups that can lead to inferior medical diagnosis and treatment,” says Yang, who along with Zhang are the two lead authors on the paper. The main point of their inquiry is to determine the kinds of subpopulation shifts that can occur and to uncover the mechanisms behind them so that, ultimately, more equitable models can be developed.

    The new paper “significantly advances our understanding” of the subpopulation shift phenomenon, claims Stanford University computer scientist Sanmi Koyejo. “This research contributes valuable insights for future advancements in machine learning models’ performance on underrepresented subgroups.”

    Camels and cattle

    The MIT group has identified four principal types of shifts — spurious correlations, attribute imbalance, class imbalance, and attribute generalization — which, according to Yang, “have never been put together into a coherent and unified framework. We’ve come up with a single equation that shows you where biases can come from.”

    Biases can, in fact, stem from what the researchers call the class, or from the attribute, or both. To pick a simple example, suppose the task assigned to the machine learning model is to sort images of objects — animals in this case — into two classes: cows and camels. Attributes are descriptors that don’t specifically relate to the class itself. It might turn out, for instance, that all the images used in the analysis show cows standing on grass and camels on sand — grass and sand serving as the attributes here. Given the data available to it, the machine could reach an erroneous conclusion — namely that cows can only be found on grass, not on sand, with the opposite being true for camels. Such a finding would be incorrect, however, giving rise to a spurious correlation, which, Yang explains, is a “special case” among subpopulation shifts — “one in which you have a bias in both the class and the attribute.”

    In a medical setting, one could rely on machine learning models to determine whether a person has pneumonia or not based on an examination of X-ray images. There would be two classes in this situation, one consisting of people who have the lung ailment, another for those who are infection-free. A relatively straightforward case would involve just two attributes: the people getting X-rayed are either female or male. If, in this particular dataset, there were 100 males diagnosed with pneumonia for every one female diagnosed with pneumonia, that could lead to an attribute imbalance, and the model would likely do a better job of correctly detecting pneumonia for a man than for a woman. Similarly, having 1,000 times more healthy (pneumonia-free) subjects than sick ones would lead to a class imbalance, with the model biased toward healthy cases. Attribute generalization is the last shift highlighted in the new study. If your sample contained 100 male patients with pneumonia and zero female subjects with the same illness, you still would like the model to be able to generalize and make predictions about female subjects even though there are no samples in the training data for females with pneumonia.

    The team then took 20 advanced algorithms, designed to carry out classification tasks, and tested them on a dozen datasets to see how they performed across different population groups. They reached some unexpected conclusions: By improving the “classifier,” which is the last layer of the neural network, they were able to reduce the occurrence of spurious correlations and class imbalance, but the other shifts were unaffected. Improvements to the “encoder,” one of the uppermost layers in the neural network, could reduce the problem of attribute imbalance. “However, no matter what we did to the encoder or classifier, we did not see any improvements in terms of attribute generalization,” Yang says, “and we don’t yet know how to address that.”

    Precisely accurate

    There is also the question of assessing how well your model actually works in terms of evenhandedness among different population groups. The metric normally used, called worst-group accuracy or WGA, is based on the assumption that if you can improve the accuracy — of, say, medical diagnosis — for the group that has the worst model performance, you would have improved the model as a whole. “The WGA is considered the gold standard in subpopulation evaluation,” the authors contend, but they made a surprising discovery: boosting worst-group accuracy results in a decrease in what they call “worst-case precision.” In medical decision-making of all sorts, one needs both accuracy — which speaks to the validity of the findings — and precision, which relates to the reliability of the methodology. “Precision and accuracy are both very important metrics in classification tasks, and that is especially true in medical diagnostics,” Yang explains. “You should never trade precision for accuracy. You always need to balance the two.”

    The MIT scientists are putting their theories into practice. In a study they’re conducting with a medical center, they’re looking at public datasets for tens of thousands of patients and hundreds of thousands of chest X-rays, trying to see whether it’s possible for machine learning models to work in an unbiased manner for all populations. That’s still far from the case, even though more awareness has been drawn to this problem, Yang says. “We are finding many disparities across different ages, gender, ethnicity, and intersectional groups.”

    He and his colleagues agree on the eventual goal, which is to achieve fairness in health care among all populations. But before we can reach that point, they maintain, we still need a better understanding of the sources of unfairness and how they permeate our current system. Reforming the system as a whole will not be easy, they acknowledge. In fact, the title of the paper they introduced at the Honolulu conference, “Change is Hard,” gives some indications as to the challenges that they and like-minded researchers face. More

  • in

    A simpler method for learning to control a robot

    Researchers from MIT and Stanford University have devised a new machine-learning approach that could be used to control a robot, such as a drone or autonomous vehicle, more effectively and efficiently in dynamic environments where conditions can change rapidly.

    This technique could help an autonomous vehicle learn to compensate for slippery road conditions to avoid going into a skid, allow a robotic free-flyer to tow different objects in space, or enable a drone to closely follow a downhill skier despite being buffeted by strong winds.

    The researchers’ approach incorporates certain structure from control theory into the process for learning a model in such a way that leads to an effective method of controlling complex dynamics, such as those caused by impacts of wind on the trajectory of a flying vehicle. One way to think about this structure is as a hint that can help guide how to control a system.

    “The focus of our work is to learn intrinsic structure in the dynamics of the system that can be leveraged to design more effective, stabilizing controllers,” says Navid Azizan, the Esther and Harold E. Edgerton Assistant Professor in the MIT Department of Mechanical Engineering and the Institute for Data, Systems, and Society (IDSS), and a member of the Laboratory for Information and Decision Systems (LIDS). “By jointly learning the system’s dynamics and these unique control-oriented structures from data, we’re able to naturally create controllers that function much more effectively in the real world.”

    Using this structure in a learned model, the researchers’ technique immediately extracts an effective controller from the model, as opposed to other machine-learning methods that require a controller to be derived or learned separately with additional steps. With this structure, their approach is also able to learn an effective controller using fewer data than other approaches. This could help their learning-based control system achieve better performance faster in rapidly changing environments.

    “This work tries to strike a balance between identifying structure in your system and just learning a model from data,” says lead author Spencer M. Richards, a graduate student at Stanford University. “Our approach is inspired by how roboticists use physics to derive simpler models for robots. Physical analysis of these models often yields a useful structure for the purposes of control — one that you might miss if you just tried to naively fit a model to data. Instead, we try to identify similarly useful structure from data that indicates how to implement your control logic.”

    Additional authors of the paper are Jean-Jacques Slotine, professor of mechanical engineering and of brain and cognitive sciences at MIT, and Marco Pavone, associate professor of aeronautics and astronautics at Stanford. The research will be presented at the International Conference on Machine Learning (ICML).

    Learning a controller

    Determining the best way to control a robot to accomplish a given task can be a difficult problem, even when researchers know how to model everything about the system.

    A controller is the logic that enables a drone to follow a desired trajectory, for example. This controller would tell the drone how to adjust its rotor forces to compensate for the effect of winds that can knock it off a stable path to reach its goal.

    This drone is a dynamical system — a physical system that evolves over time. In this case, its position and velocity change as it flies through the environment. If such a system is simple enough, engineers can derive a controller by hand. 

    Modeling a system by hand intrinsically captures a certain structure based on the physics of the system. For instance, if a robot were modeled manually using differential equations, these would capture the relationship between velocity, acceleration, and force. Acceleration is the rate of change in velocity over time, which is determined by the mass of and forces applied to the robot.

    But often the system is too complex to be exactly modeled by hand. Aerodynamic effects, like the way swirling wind pushes a flying vehicle, are notoriously difficult to derive manually, Richards explains. Researchers would instead take measurements of the drone’s position, velocity, and rotor speeds over time, and use machine learning to fit a model of this dynamical system to the data. But these approaches typically don’t learn a control-based structure. This structure is useful in determining how to best set the rotor speeds to direct the motion of the drone over time.

    Once they have modeled the dynamical system, many existing approaches also use data to learn a separate controller for the system.

    “Other approaches that try to learn dynamics and a controller from data as separate entities are a bit detached philosophically from the way we normally do it for simpler systems. Our approach is more reminiscent of deriving models by hand from physics and linking that to control,” Richards says.

    Identifying structure

    The team from MIT and Stanford developed a technique that uses machine learning to learn the dynamics model, but in such a way that the model has some prescribed structure that is useful for controlling the system.

    With this structure, they can extract a controller directly from the dynamics model, rather than using data to learn an entirely separate model for the controller.

    “We found that beyond learning the dynamics, it’s also essential to learn the control-oriented structure that supports effective controller design. Our approach of learning state-dependent coefficient factorizations of the dynamics has outperformed the baselines in terms of data efficiency and tracking capability, proving to be successful in efficiently and effectively controlling the system’s trajectory,” Azizan says. 

    When they tested this approach, their controller closely followed desired trajectories, outpacing all the baseline methods. The controller extracted from their learned model nearly matched the performance of a ground-truth controller, which is built using the exact dynamics of the system.

    “By making simpler assumptions, we got something that actually worked better than other complicated baseline approaches,” Richards adds.

    The researchers also found that their method was data-efficient, which means it achieved high performance even with few data. For instance, it could effectively model a highly dynamic rotor-driven vehicle using only 100 data points. Methods that used multiple learned components saw their performance drop much faster with smaller datasets.

    This efficiency could make their technique especially useful in situations where a drone or robot needs to learn quickly in rapidly changing conditions.

    Plus, their approach is general and could be applied to many types of dynamical systems, from robotic arms to free-flying spacecraft operating in low-gravity environments.

    In the future, the researchers are interested in developing models that are more physically interpretable, and that would be able to identify very specific information about a dynamical system, Richards says. This could lead to better-performing controllers.

    “Despite its ubiquity and importance, nonlinear feedback control remains an art, making it especially suitable for data-driven and learning-based methods. This paper makes a significant contribution to this area by proposing a method that jointly learns system dynamics, a controller, and control-oriented structure,” says Nikolai Matni, an assistant professor in the Department of Electrical and Systems Engineering at the University of Pennsylvania, who was not involved with this work. “What I found particularly exciting and compelling was the integration of these components into a joint learning algorithm, such that control-oriented structure acts as an inductive bias in the learning process. The result is a data-efficient learning process that outputs dynamic models that enjoy intrinsic structure that enables effective, stable, and robust control. While the technical contributions of the paper are excellent themselves, it is this conceptual contribution that I view as most exciting and significant.”

    This research is supported, in part, by the NASA University Leadership Initiative and the Natural Sciences and Engineering Research Council of Canada. More

  • in

    A faster way to teach a robot

    Imagine purchasing a robot to perform household tasks. This robot was built and trained in a factory on a certain set of tasks and has never seen the items in your home. When you ask it to pick up a mug from your kitchen table, it might not recognize your mug (perhaps because this mug is painted with an unusual image, say, of MIT’s mascot, Tim the Beaver). So, the robot fails.

    “Right now, the way we train these robots, when they fail, we don’t really know why. So you would just throw up your hands and say, ‘OK, I guess we have to start over.’ A critical component that is missing from this system is enabling the robot to demonstrate why it is failing so the user can give it feedback,” says Andi Peng, an electrical engineering and computer science (EECS) graduate student at MIT.

    Peng and her collaborators at MIT, New York University, and the University of California at Berkeley created a framework that enables humans to quickly teach a robot what they want it to do, with a minimal amount of effort.

    When a robot fails, the system uses an algorithm to generate counterfactual explanations that describe what needed to change for the robot to succeed. For instance, maybe the robot would have been able to pick up the mug if the mug were a certain color. It shows these counterfactuals to the human and asks for feedback on why the robot failed. Then the system utilizes this feedback and the counterfactual explanations to generate new data it uses to fine-tune the robot.

    Fine-tuning involves tweaking a machine-learning model that has already been trained to perform one task, so it can perform a second, similar task.

    The researchers tested this technique in simulations and found that it could teach a robot more efficiently than other methods. The robots trained with this framework performed better, while the training process consumed less of a human’s time.

    This framework could help robots learn faster in new environments without requiring a user to have technical knowledge. In the long run, this could be a step toward enabling general-purpose robots to efficiently perform daily tasks for the elderly or individuals with disabilities in a variety of settings.

    Peng, the lead author, is joined by co-authors Aviv Netanyahu, an EECS graduate student; Mark Ho, an assistant professor at the Stevens Institute of Technology; Tianmin Shu, an MIT postdoc; Andreea Bobu, a graduate student at UC Berkeley; and senior authors Julie Shah, an MIT professor of aeronautics and astronautics and the director of the Interactive Robotics Group in the Computer Science and Artificial Intelligence Laboratory (CSAIL), and Pulkit Agrawal, a professor in CSAIL. The research will be presented at the International Conference on Machine Learning.

    On-the-job training

    Robots often fail due to distribution shift — the robot is presented with objects and spaces it did not see during training, and it doesn’t understand what to do in this new environment.

    One way to retrain a robot for a specific task is imitation learning. The user could demonstrate the correct task to teach the robot what to do. If a user tries to teach a robot to pick up a mug, but demonstrates with a white mug, the robot could learn that all mugs are white. It may then fail to pick up a red, blue, or “Tim-the-Beaver-brown” mug.

    Training a robot to recognize that a mug is a mug, regardless of its color, could take thousands of demonstrations.

    “I don’t want to have to demonstrate with 30,000 mugs. I want to demonstrate with just one mug. But then I need to teach the robot so it recognizes that it can pick up a mug of any color,” Peng says.

    To accomplish this, the researchers’ system determines what specific object the user cares about (a mug) and what elements aren’t important for the task (perhaps the color of the mug doesn’t matter). It uses this information to generate new, synthetic data by changing these “unimportant” visual concepts. This process is known as data augmentation.

    The framework has three steps. First, it shows the task that caused the robot to fail. Then it collects a demonstration from the user of the desired actions and generates counterfactuals by searching over all features in the space that show what needed to change for the robot to succeed.

    The system shows these counterfactuals to the user and asks for feedback to determine which visual concepts do not impact the desired action. Then it uses this human feedback to generate many new augmented demonstrations.

    In this way, the user could demonstrate picking up one mug, but the system would produce demonstrations showing the desired action with thousands of different mugs by altering the color. It uses these data to fine-tune the robot.

    Creating counterfactual explanations and soliciting feedback from the user are critical for the technique to succeed, Peng says.

    From human reasoning to robot reasoning

    Because their work seeks to put the human in the training loop, the researchers tested their technique with human users. They first conducted a study in which they asked people if counterfactual explanations helped them identify elements that could be changed without affecting the task.

    “It was so clear right off the bat. Humans are so good at this type of counterfactual reasoning. And this counterfactual step is what allows human reasoning to be translated into robot reasoning in a way that makes sense,” she says.

    Then they applied their framework to three simulations where robots were tasked with: navigating to a goal object, picking up a key and unlocking a door, and picking up a desired object then placing it on a tabletop. In each instance, their method enabled the robot to learn faster than with other techniques, while requiring fewer demonstrations from users.

    Moving forward, the researchers hope to test this framework on real robots. They also want to focus on reducing the time it takes the system to create new data using generative machine-learning models.

    “We want robots to do what humans do, and we want them to do it in a semantically meaningful way. Humans tend to operate in this abstract space, where they don’t think about every single property in an image. At the end of the day, this is really about enabling a robot to learn a good, human-like representation at an abstract level,” Peng says.

    This research is supported, in part, by a National Science Foundation Graduate Research Fellowship, Open Philanthropy, an Apple AI/ML Fellowship, Hyundai Motor Corporation, the MIT-IBM Watson AI Lab, and the National Science Foundation Institute for Artificial Intelligence and Fundamental Interactions. More

  • in

    A new way to look at data privacy

    Imagine that a team of scientists has developed a machine-learning model that can predict whether a patient has cancer from lung scan images. They want to share this model with hospitals around the world so clinicians can start using it in diagnosis.

    But there’s a problem. To teach their model how to predict cancer, they showed it millions of real lung scan images, a process called training. Those sensitive data, which are now encoded into the inner workings of the model, could potentially be extracted by a malicious agent. The scientists can prevent this by adding noise, or more generic randomness, to the model that makes it harder for an adversary to guess the original data. However, perturbation reduces a model’s accuracy, so the less noise one can add, the better.

    MIT researchers have developed a technique that enables the user to potentially add the smallest amount of noise possible, while still ensuring the sensitive data are protected.

    The researchers created a new privacy metric, which they call Probably Approximately Correct (PAC) Privacy, and built a framework based on this metric that can automatically determine the minimal amount of noise that needs to be added. Moreover, this framework does not need knowledge of the inner workings of a model or its training process, which makes it easier to use for different types of models and applications.

    In several cases, the researchers show that the amount of noise required to protect sensitive data from adversaries is far less with PAC Privacy than with other approaches. This could help engineers create machine-learning models that provably hide training data, while maintaining accuracy in real-world settings.

    “PAC Privacy exploits the uncertainty or entropy of the sensitive data in a meaningful way,  and this allows us to add, in many cases, an order of magnitude less noise. This framework allows us to understand the characteristics of arbitrary data processing and privatize it automatically without artificial modifications. While we are in the early days and we are doing simple examples, we are excited about the promise of this technique,” says Srini Devadas, the Edwin Sibley Webster Professor of Electrical Engineering and co-author of a new paper on PAC Privacy.

    Devadas wrote the paper with lead author Hanshen Xiao, an electrical engineering and computer science graduate student. The research will be presented at the International Cryptography Conference (Crypto 2023).

    Defining privacy

    A fundamental question in data privacy is: How much sensitive data could an adversary recover from a machine-learning model with noise added to it?

    Differential Privacy, one popular privacy definition, says privacy is achieved if an adversary who observes the released model cannot infer whether an arbitrary individual’s data is used for the training processing. But provably preventing an adversary from distinguishing data usage often requires large amounts of noise to obscure it. This noise reduces the model’s accuracy.

    PAC Privacy looks at the problem a bit differently. It characterizes how hard it would be for an adversary to reconstruct any part of randomly sampled or generated sensitive data after noise has been added, rather than only focusing on the distinguishability problem.

    For instance, if the sensitive data are images of human faces, differential privacy would focus on whether the adversary can tell if someone’s face was in the dataset. PAC Privacy, on the other hand, could look at whether an adversary could extract a silhouette — an approximation — that someone could recognize as a particular individual’s face.

    Once they established the definition of PAC Privacy, the researchers created an algorithm that automatically tells the user how much noise to add to a model to prevent an adversary from confidently reconstructing a close approximation of the sensitive data. This algorithm guarantees privacy even if the adversary has infinite computing power, Xiao says.

    To find the optimal amount of noise, the PAC Privacy algorithm relies on the uncertainty, or entropy, in the original data from the viewpoint of the adversary.

    This automatic technique takes samples randomly from a data distribution or a large data pool and runs the user’s machine-learning training algorithm on that subsampled data to produce an output learned model. It does this many times on different subsamplings and compares the variance across all outputs. This variance determines how much noise one must add — a smaller variance means less noise is needed.

    Algorithm advantages

    Different from other privacy approaches, the PAC Privacy algorithm does not need knowledge of the inner workings of a model, or the training process.

    When implementing PAC Privacy, a user can specify their desired level of confidence at the outset. For instance, perhaps the user wants a guarantee that an adversary will not be more than 1 percent confident that they have successfully reconstructed the sensitive data to within 5 percent of its actual value. The PAC Privacy algorithm automatically tells the user the optimal amount of noise that needs to be added to the output model before it is shared publicly, in order to achieve those goals.

    “The noise is optimal, in the sense that if you add less than we tell you, all bets could be off. But the effect of adding noise to neural network parameters is complicated, and we are making no promises on the utility drop the model may experience with the added noise,” Xiao says.

    This points to one limitation of PAC Privacy — the technique does not tell the user how much accuracy the model will lose once the noise is added. PAC Privacy also involves repeatedly training a machine-learning model on many subsamplings of data, so it can be computationally expensive.  

    To improve PAC Privacy, one approach is to modify a user’s machine-learning training process so it is more stable, meaning that the output model it produces does not change very much when the input data is subsampled from a data pool.  This stability would create smaller variances between subsample outputs, so not only would the PAC Privacy algorithm need to be run fewer times to identify the optimal amount of noise, but it would also need to add less noise.

    An added benefit of stabler models is that they often have less generalization error, which means they can make more accurate predictions on previously unseen data, a win-win situation between machine learning and privacy, Devadas adds.

    “In the next few years, we would love to look a little deeper into this relationship between stability and privacy, and the relationship between privacy and generalization error. We are knocking on a door here, but it is not clear yet where the door leads,” he says.

    “Obfuscating the usage of an individual’s data in a model is paramount to protecting their privacy. However, to do so can come at the cost of the datas’ and therefore model’s utility,” says Jeremy Goodsitt, senior machine learning engineer at Capital One, who was not involved with this research. “PAC provides an empirical, black-box solution, which can reduce the added noise compared to current practices while maintaining equivalent privacy guarantees. In addition, its empirical approach broadens its reach to more data consuming applications.”

    This research is funded, in part, by DSTA Singapore, Cisco Systems, Capital One, and a MathWorks Fellowship. More

  • in

    Learning the language of molecules to predict their properties

    Discovering new materials and drugs typically involves a manual, trial-and-error process that can take decades and cost millions of dollars. To streamline this process, scientists often use machine learning to predict molecular properties and narrow down the molecules they need to synthesize and test in the lab.

    Researchers from MIT and the MIT-Watson AI Lab have developed a new, unified framework that can simultaneously predict molecular properties and generate new molecules much more efficiently than these popular deep-learning approaches.

    To teach a machine-learning model to predict a molecule’s biological or mechanical properties, researchers must show it millions of labeled molecular structures — a process known as training. Due to the expense of discovering molecules and the challenges of hand-labeling millions of structures, large training datasets are often hard to come by, which limits the effectiveness of machine-learning approaches.

    By contrast, the system created by the MIT researchers can effectively predict molecular properties using only a small amount of data. Their system has an underlying understanding of the rules that dictate how building blocks combine to produce valid molecules. These rules capture the similarities between molecular structures, which helps the system generate new molecules and predict their properties in a data-efficient manner.

    This method outperformed other machine-learning approaches on both small and large datasets, and was able to accurately predict molecular properties and generate viable molecules when given a dataset with fewer than 100 samples.

    “Our goal with this project is to use some data-driven methods to speed up the discovery of new molecules, so you can train a model to do the prediction without all of these cost-heavy experiments,” says lead author Minghao Guo, a computer science and electrical engineering (EECS) graduate student.

    Guo’s co-authors include MIT-IBM Watson AI Lab research staff members Veronika Thost, Payel Das, and Jie Chen; recent MIT graduates Samuel Song ’23 and Adithya Balachandran ’23; and senior author Wojciech Matusik, a professor of electrical engineering and computer science and a member of the MIT-IBM Watson AI Lab, who leads the Computational Design and Fabrication Group within the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL). The research will be presented at the International Conference for Machine Learning.

    Learning the language of molecules

    To achieve the best results with machine-learning models, scientists need training datasets with millions of molecules that have similar properties to those they hope to discover. In reality, these domain-specific datasets are usually very small. So, researchers use models that have been pretrained on large datasets of general molecules, which they apply to a much smaller, targeted dataset. However, because these models haven’t acquired much domain-specific knowledge, they tend to perform poorly.

    The MIT team took a different approach. They created a machine-learning system that automatically learns the “language” of molecules — what is known as a molecular grammar — using only a small, domain-specific dataset. It uses this grammar to construct viable molecules and predict their properties.

    In language theory, one generates words, sentences, or paragraphs based on a set of grammar rules. You can think of a molecular grammar the same way. It is a set of production rules that dictate how to generate molecules or polymers by combining atoms and substructures.

    Just like a language grammar, which can generate a plethora of sentences using the same rules, one molecular grammar can represent a vast number of molecules. Molecules with similar structures use the same grammar production rules, and the system learns to understand these similarities.

    Since structurally similar molecules often have similar properties, the system uses its underlying knowledge of molecular similarity to predict properties of new molecules more efficiently. 

    “Once we have this grammar as a representation for all the different molecules, we can use it to boost the process of property prediction,” Guo says.

    The system learns the production rules for a molecular grammar using reinforcement learning — a trial-and-error process where the model is rewarded for behavior that gets it closer to achieving a goal.

    But because there could be billions of ways to combine atoms and substructures, the process to learn grammar production rules would be too computationally expensive for anything but the tiniest dataset.

    The researchers decoupled the molecular grammar into two parts. The first part, called a metagrammar, is a general, widely applicable grammar they design manually and give the system at the outset. Then it only needs to learn a much smaller, molecule-specific grammar from the domain dataset. This hierarchical approach speeds up the learning process.

    Big results, small datasets

    In experiments, the researchers’ new system simultaneously generated viable molecules and polymers, and predicted their properties more accurately than several popular machine-learning approaches, even when the domain-specific datasets had only a few hundred samples. Some other methods also required a costly pretraining step that the new system avoids.

    The technique was especially effective at predicting physical properties of polymers, such as the glass transition temperature, which is the temperature required for a material to transition from solid to liquid. Obtaining this information manually is often extremely costly because the experiments require extremely high temperatures and pressures.

    To push their approach further, the researchers cut one training set down by more than half — to just 94 samples. Their model still achieved results that were on par with methods trained using the entire dataset.

    “This grammar-based representation is very powerful. And because the grammar itself is a very general representation, it can be deployed to different kinds of graph-form data. We are trying to identify other applications beyond chemistry or material science,” Guo says.

    In the future, they also want to extend their current molecular grammar to include the 3D geometry of molecules and polymers, which is key to understanding the interactions between polymer chains. They are also developing an interface that would show a user the learned grammar production rules and solicit feedback to correct rules that may be wrong, boosting the accuracy of the system.

    This work is funded, in part, by the MIT-IBM Watson AI Lab and its member company, Evonik. More