More stories

  • in

    AI that can learn the patterns of human language

    Human languages are notoriously complex, and linguists have long thought it would be impossible to teach a machine how to analyze speech sounds and word structures in the way human investigators do.

    But researchers at MIT, Cornell University, and McGill University have taken a step in this direction. They have demonstrated an artificial intelligence system that can learn the rules and patterns of human languages on its own.

    When given words and examples of how those words change to express different grammatical functions (like tense, case, or gender) in one language, this machine-learning model comes up with rules that explain why the forms of those words change. For instance, it might learn that the letter “a” must be added to end of a word to make the masculine form feminine in Serbo-Croatian.

    This model can also automatically learn higher-level language patterns that can apply to many languages, enabling it to achieve better results.

    The researchers trained and tested the model using problems from linguistics textbooks that featured 58 different languages. Each problem had a set of words and corresponding word-form changes. The model was able to come up with a correct set of rules to describe those word-form changes for 60 percent of the problems.

    This system could be used to study language hypotheses and investigate subtle similarities in the way diverse languages transform words. It is especially unique because the system discovers models that can be readily understood by humans, and it acquires these models from small amounts of data, such as a few dozen words. And instead of using one massive dataset for a single task, the system utilizes many small datasets, which is closer to how scientists propose hypotheses — they look at multiple related datasets and come up with models to explain phenomena across those datasets.

    “One of the motivations of this work was our desire to study systems that learn models of datasets that is represented in a way that humans can understand. Instead of learning weights, can the model learn expressions or rules? And we wanted to see if we could build this system so it would learn on a whole battery of interrelated datasets, to make the system learn a little bit about how to better model each one,” says Kevin Ellis ’14, PhD ’20, an assistant professor of computer science at Cornell University and lead author of the paper.

    Joining Ellis on the paper are MIT faculty members Adam Albright, a professor of linguistics; Armando Solar-Lezama, a professor and associate director of the Computer Science and Artificial Intelligence Laboratory (CSAIL); and Joshua B. Tenenbaum, the Paul E. Newton Career Development Professor of Cognitive Science and Computation in the Department of Brain and Cognitive Sciences and a member of CSAIL; as well as senior author

    Timothy J. O’Donnell, assistant professor in the Department of Linguistics at McGill University, and Canada CIFAR AI Chair at the Mila – Quebec Artificial Intelligence Institute.

    The research is published today in Nature Communications.

    Looking at language 

    In their quest to develop an AI system that could automatically learn a model from multiple related datasets, the researchers chose to explore the interaction of phonology (the study of sound patterns) and morphology (the study of word structure).

    Data from linguistics textbooks offered an ideal testbed because many languages share core features, and textbook problems showcase specific linguistic phenomena. Textbook problems can also be solved by college students in a fairly straightforward way, but those students typically have prior knowledge about phonology from past lessons they use to reason about new problems.

    Ellis, who earned his PhD at MIT and was jointly advised by Tenenbaum and Solar-Lezama, first learned about morphology and phonology in an MIT class co-taught by O’Donnell, who was a postdoc at the time, and Albright.

    “Linguists have thought that in order to really understand the rules of a human language, to empathize with what it is that makes the system tick, you have to be human. We wanted to see if we can emulate the kinds of knowledge and reasoning that humans (linguists) bring to the task,” says Albright.

    To build a model that could learn a set of rules for assembling words, which is called a grammar, the researchers used a machine-learning technique known as Bayesian Program Learning. With this technique, the model solves a problem by writing a computer program.

    In this case, the program is the grammar the model thinks is the most likely explanation of the words and meanings in a linguistics problem. They built the model using Sketch, a popular program synthesizer which was developed at MIT by Solar-Lezama.

    But Sketch can take a lot of time to reason about the most likely program. To get around this, the researchers had the model work one piece at a time, writing a small program to explain some data, then writing a larger program that modifies that small program to cover more data, and so on.

    They also designed the model so it learns what “good” programs tend to look like. For instance, it might learn some general rules on simple Russian problems that it would apply to a more complex problem in Polish because the languages are similar. This makes it easier for the model to solve the Polish problem.

    Tackling textbook problems

    When they tested the model using 70 textbook problems, it was able to find a grammar that matched the entire set of words in the problem in 60 percent of cases, and correctly matched most of the word-form changes in 79 percent of problems.

    The researchers also tried pre-programming the model with some knowledge it “should” have learned if it was taking a linguistics course, and showed that it could solve all problems better.

    “One challenge of this work was figuring out whether what the model was doing was reasonable. This isn’t a situation where there is one number that is the single right answer. There is a range of possible solutions which you might accept as right, close to right, etc.,” Albright says.

    The model often came up with unexpected solutions. In one instance, it discovered the expected answer to a Polish language problem, but also another correct answer that exploited a mistake in the textbook. This shows that the model could “debug” linguistics analyses, Ellis says.

    The researchers also conducted tests that showed the model was able to learn some general templates of phonological rules that could be applied across all problems.

    “One of the things that was most surprising is that we could learn across languages, but it didn’t seem to make a huge difference,” says Ellis. “That suggests two things. Maybe we need better methods for learning across problems. And maybe, if we can’t come up with those methods, this work can help us probe different ideas we have about what knowledge to share across problems.”

    In the future, the researchers want to use their model to find unexpected solutions to problems in other domains. They could also apply the technique to more situations where higher-level knowledge can be applied across interrelated datasets. For instance, perhaps they could develop a system to infer differential equations from datasets on the motion of different objects, says Ellis.

    “This work shows that we have some methods which can, to some extent, learn inductive biases. But I don’t think we’ve quite figured out, even for these textbook problems, the inductive bias that lets a linguist accept the plausible grammars and reject the ridiculous ones,” he adds.

    “This work opens up many exciting venues for future research. I am particularly intrigued by the possibility that the approach explored by Ellis and colleagues (Bayesian Program Learning, BPL) might speak to how infants acquire language,” says T. Florian Jaeger, a professor of brain and cognitive sciences and computer science at the University of Rochester, who was not an author of this paper. “Future work might ask, for example, under what additional induction biases (assumptions about universal grammar) the BPL approach can successfully achieve human-like learning behavior on the type of data infants observe during language acquisition. I think it would be fascinating to see whether inductive biases that are even more abstract than those considered by Ellis and his team — such as biases originating in the limits of human information processing (e.g., memory constraints on dependency length or capacity limits in the amount of information that can be processed per time) — would be sufficient to induce some of the patterns observed in human languages.”

    This work was funded, in part, by the Air Force Office of Scientific Research, the Center for Brains, Minds, and Machines, the MIT-IBM Watson AI Lab, the Natural Science and Engineering Research Council of Canada, the Fonds de Recherche du Québec – Société et Culture, the Canada CIFAR AI Chairs Program, the National Science Foundation (NSF), and an NSF graduate fellowship. More

  • in

    Taking a magnifying glass to data center operations

    When the MIT Lincoln Laboratory Supercomputing Center (LLSC) unveiled its TX-GAIA supercomputer in 2019, it provided the MIT community a powerful new resource for applying artificial intelligence to their research. Anyone at MIT can submit a job to the system, which churns through trillions of operations per second to train models for diverse applications, such as spotting tumors in medical images, discovering new drugs, or modeling climate effects. But with this great power comes the great responsibility of managing and operating it in a sustainable manner — and the team is looking for ways to improve.

    “We have these powerful computational tools that let researchers build intricate models to solve problems, but they can essentially be used as black boxes. What gets lost in there is whether we are actually using the hardware as effectively as we can,” says Siddharth Samsi, a research scientist in the LLSC. 

    To gain insight into this challenge, the LLSC has been collecting detailed data on TX-GAIA usage over the past year. More than a million user jobs later, the team has released the dataset open source to the computing community.

    Their goal is to empower computer scientists and data center operators to better understand avenues for data center optimization — an important task as processing needs continue to grow. They also see potential for leveraging AI in the data center itself, by using the data to develop models for predicting failure points, optimizing job scheduling, and improving energy efficiency. While cloud providers are actively working on optimizing their data centers, they do not often make their data or models available for the broader high-performance computing (HPC) community to leverage. The release of this dataset and associated code seeks to fill this space.

    “Data centers are changing. We have an explosion of hardware platforms, the types of workloads are evolving, and the types of people who are using data centers is changing,” says Vijay Gadepally, a senior researcher at the LLSC. “Until now, there hasn’t been a great way to analyze the impact to data centers. We see this research and dataset as a big step toward coming up with a principled approach to understanding how these variables interact with each other and then applying AI for insights and improvements.”

    Papers describing the dataset and potential applications have been accepted to a number of venues, including the IEEE International Symposium on High-Performance Computer Architecture, the IEEE International Parallel and Distributed Processing Symposium, the Annual Conference of the North American Chapter of the Association for Computational Linguistics, the IEEE High-Performance and Embedded Computing Conference, and International Conference for High Performance Computing, Networking, Storage and Analysis. 

    Workload classification

    Among the world’s TOP500 supercomputers, TX-GAIA combines traditional computing hardware (central processing units, or CPUs) with nearly 900 graphics processing unit (GPU) accelerators. These NVIDIA GPUs are specialized for deep learning, the class of AI that has given rise to speech recognition and computer vision.

    The dataset covers CPU, GPU, and memory usage by job; scheduling logs; and physical monitoring data. Compared to similar datasets, such as those from Google and Microsoft, the LLSC dataset offers “labeled data, a variety of known AI workloads, and more detailed time series data compared with prior datasets. To our knowledge, it’s one of the most comprehensive and fine-grained datasets available,” Gadepally says. 

    Notably, the team collected time-series data at an unprecedented level of detail: 100-millisecond intervals on every GPU and 10-second intervals on every CPU, as the machines processed more than 3,000 known deep-learning jobs. One of the first goals is to use this labeled dataset to characterize the workloads that different types of deep-learning jobs place on the system. This process would extract features that reveal differences in how the hardware processes natural language models versus image classification or materials design models, for example.   

    The team has now launched the MIT Datacenter Challenge to mobilize this research. The challenge invites researchers to use AI techniques to identify with 95 percent accuracy the type of job that was run, using their labeled time-series data as ground truth.

    Such insights could enable data centers to better match a user’s job request with the hardware best suited for it, potentially conserving energy and improving system performance. Classifying workloads could also allow operators to quickly notice discrepancies resulting from hardware failures, inefficient data access patterns, or unauthorized usage.

    Too many choices

    Today, the LLSC offers tools that let users submit their job and select the processors they want to use, “but it’s a lot of guesswork on the part of users,” Samsi says. “Somebody might want to use the latest GPU, but maybe their computation doesn’t actually need it and they could get just as impressive results on CPUs, or lower-powered machines.”

    Professor Devesh Tiwari at Northeastern University is working with the LLSC team to develop techniques that can help users match their workloads to appropriate hardware. Tiwari explains that the emergence of different types of AI accelerators, GPUs, and CPUs has left users suffering from too many choices. Without the right tools to take advantage of this heterogeneity, they are missing out on the benefits: better performance, lower costs, and greater productivity.

    “We are fixing this very capability gap — making users more productive and helping users do science better and faster without worrying about managing heterogeneous hardware,” says Tiwari. “My PhD student, Baolin Li, is building new capabilities and tools to help HPC users leverage heterogeneity near-optimally without user intervention, using techniques grounded in Bayesian optimization and other learning-based optimization methods. But, this is just the beginning. We are looking into ways to introduce heterogeneity in our data centers in a principled approach to help our users achieve the maximum advantage of heterogeneity autonomously and cost-effectively.”

    Workload classification is the first of many problems to be posed through the Datacenter Challenge. Others include developing AI techniques to predict job failures, conserve energy, or create job scheduling approaches that improve data center cooling efficiencies.

    Energy conservation 

    To mobilize research into greener computing, the team is also planning to release an environmental dataset of TX-GAIA operations, containing rack temperature, power consumption, and other relevant data.

    According to the researchers, huge opportunities exist to improve the power efficiency of HPC systems being used for AI processing. As one example, recent work in the LLSC determined that simple hardware tuning, such as limiting the amount of power an individual GPU can draw, could reduce the energy cost of training an AI model by 20 percent, with only modest increases in computing time. “This reduction translates to approximately an entire week’s worth of household energy for a mere three-hour time increase,” Gadepally says.

    They have also been developing techniques to predict model accuracy, so that users can quickly terminate experiments that are unlikely to yield meaningful results, saving energy. The Datacenter Challenge will share relevant data to enable researchers to explore other opportunities to conserve energy.

    The team expects that lessons learned from this research can be applied to the thousands of data centers operated by the U.S. Department of Defense. The U.S. Air Force is a sponsor of this work, which is being conducted under the USAF-MIT AI Accelerator.

    Other collaborators include researchers at MIT Computer Science and Artificial Intelligence Laboratory (CSAIL). Professor Charles Leiserson’s Supertech Research Group is investigating performance-enhancing techniques for parallel computing, and research scientist Neil Thompson is designing studies on ways to nudge data center users toward climate-friendly behavior.

    Samsi presented this work at the inaugural AI for Datacenter Optimization (ADOPT’22) workshop last spring as part of the IEEE International Parallel and Distributed Processing Symposium. The workshop officially introduced their Datacenter Challenge to the HPC community.

    “We hope this research will allow us and others who run supercomputing centers to be more responsive to user needs while also reducing the energy consumption at the center level,” Samsi says. More

  • in

    Researchers discover major roadblock in alleviating network congestion

    When users want to send data over the internet faster than the network can handle, congestion can occur — the same way traffic congestion snarls the morning commute into a big city.

    Computers and devices that transmit data over the internet break the data down into smaller packets and use a special algorithm to decide how fast to send those packets. These congestion control algorithms seek to fully discover and utilize available network capacity while sharing it fairly with other users who may be sharing the same network. These algorithms try to minimize delay caused by data waiting in queues in the network.

    Over the past decade, researchers in industry and academia have developed several algorithms that attempt to achieve high rates while controlling delays. Some of these, such as the BBR algorithm developed by Google, are now widely used by many websites and applications.

    But a team of MIT researchers has discovered that these algorithms can be deeply unfair. In a new study, they show there will always be a network scenario where at least one sender receives almost no bandwidth compared to other senders; that is, a problem known as “starvation” cannot be avoided.

    “What is really surprising about this paper and the results is that when you take into account the real-world complexity of network paths and all the things they can do to data packets, it is basically impossible for delay-controlling congestion control algorithms to avoid starvation using current methods,” says Mohammad Alizadeh, associate professor of electrical engineering and computer science (EECS).

    While Alizadeh and his co-authors weren’t able to find a traditional congestion control algorithm that could avoid starvation, there may be algorithms in a different class that could prevent this problem. Their analysis also suggests that changing how these algorithms work, so that they allow for larger variations in delay, could help prevent starvation in some network situations.

    Alizadeh wrote the paper with first author and EECS graduate student Venkat Arun and senior author Hari Balakrishnan, the Fujitsu Professor of Computer Science and Artificial Intelligence. The research will be presented at the ACM Special Interest Group on Data Communications (SIGCOMM) conference.

    Controlling congestion

    Congestion control is a fundamental problem in networking that researchers have been trying to tackle since the 1980s.

    A user’s computer does not know how fast to send data packets over the network because it lacks information, such as the quality of the network connection or how many other senders are using the network. Sending packets too slowly makes poor use of the available bandwidth. But sending them too quickly can overwhelm the network, and in doing so, packets will start to get dropped. These packets must be resent, which leads to longer delays. Delays can also be caused by packets waiting in queues for a long time.

    Congestion control algorithms use packet losses and delays as signals to infer congestion and decide how fast to send data. But the internet is complicated, and packets can be delayed and lost for reasons unrelated to network congestion. For instance, data could be held up in a queue along the way and then released with a burst of other packets, or the receiver’s acknowledgement might be delayed. The authors call delays that are not caused by congestion “jitter.”

    Even if a congestion control algorithm measures delay perfectly, it can’t tell the difference between delay caused by congestion and delay caused by jitter. Delay caused by jitter is unpredictable and confuses the sender. Because of this ambiguity, users start estimating delay differently, which causes them to send packets at unequal rates. Eventually, this leads to a situation where starvation occurs and someone gets shut out completely, Arun explains.

    “We started the project because we lacked a theoretical understanding of congestion control behavior in the presence of jitter. To place it on a firmer theoretical footing, we built a mathematical model that was simple enough to think about, yet able to capture some of the complexities of the internet. It has been very rewarding to have math tell us things we didn’t know and that have practical relevance,” he says.

    Studying starvation

    The researchers fed their mathematical model to a computer, gave it a series of commonly used congestion control algorithms, and asked the computer to find an algorithm that could avoid starvation, using their model.

    “We couldn’t do it. We tried every algorithm that we are aware of, and some new ones we made up. Nothing worked. The computer always found a situation where some people get all the bandwidth and at least one person gets basically nothing,” Arun says.

    The researchers were surprised by this result, especially since these algorithms are widely believed to be reasonably fair. They started suspecting that it may not be possible to avoid starvation, an extreme form of unfairness. This motivated them to define a class of algorithms they call “delay-convergent algorithms” that they proved will always suffer from starvation under their network model. All existing congestion control algorithms that control delay (that the researchers are aware of) are delay-convergent.

    The fact that such simple failure modes of these widely used algorithms remained unknown for so long illustrates how difficult it is to understand algorithms through empirical testing alone, Arun adds. It underscores the importance of a solid theoretical foundation.

    But all hope is not lost. While all the algorithms they tested failed, there may be other algorithms which are not delay-convergent that might be able to avoid starvation This suggests that one way to fix the problem might be to design congestion control algorithms that vary the delay range more widely, so the range is larger than any delay that might occur due to jitter in the network.

    “To control delays, algorithms have tried to also bound the variations in delay about a desired equilibrium, but there is nothing wrong in potentially creating greater delay variation to get better measurements of congestive delays. It is just a new design philosophy you would have to adopt,” Balakrishnan adds.

    Now, the researchers want to keep pushing to see if they can find or build an algorithm that will eliminate starvation. They also want to apply this approach of mathematical modeling and computational proofs to other thorny, unsolved problems in networked systems.

    “We are increasingly reliant on computer systems for very critical things, and we need to put their reliability on a firmer conceptual footing. We’ve shown the surprising things you can discover when you put in the time to come up with these formal specifications of what the problem actually is,” says Alizadeh.

    The NASA University Leadership Initiative (grant #80NSSC20M0163) provided funds to assist the authors with their research, but the research paper solely reflects the opinions and conclusions of its authors and not any NASA entity. This work was also partially funded by the National Science Foundation, award number 1751009. More

  • in

    Costis Daskalakis appointed inaugural Avanessians Professor in the MIT Schwarzman College of Computing

    The MIT Stephen A. Schwarzman College of Computing has named Costis Daskalakis as the inaugural holder of the Avanessians Professorship. His chair began on July 1.

    Daskalakis is the first person appointed to this position generously endowed by Armen Avanessians ’81. Established in the MIT Schwarzman College of Computing, the new chair provides Daskalakis with additional support to pursue his research and develop his career.

    “I’m delighted to recognize Costis for his scholarship and extraordinary achievements with this distinguished professorship,” says Daniel Huttenlocher, dean of the MIT Schwarzman College of Computing and the Henry Ellis Warren Professor of Electrical Engineering and Computer Science.

    A professor in the MIT Department of Electrical Engineering and Computer Science, Daskalakis is a theoretical computer scientist who works at the interface of game theory, economics, probability theory, statistics, and machine learning. He has resolved long-standing open problems about the computational complexity of the Nash equilibrium, the mathematical structure and computational complexity of multi-item auctions, and the behavior of machine-learning methods such as the expectation-maximization algorithm. He has obtained computationally and statistically efficient methods for statistical hypothesis testing and learning in high-dimensional settings, as well as results characterizing the structure and concentration properties of high-dimensional distributions. His current work focuses on multi-agent learning, learning from biased and dependent data, causal inference, and econometrics.

    A native of Greece, Daskalakis joined the MIT faculty in 2009. He is a member of the Computer Science and Artificial Intelligence Laboratory and is affiliated with the Laboratory for Information and Decision Systems and the Operations Research Center. He is also an investigator in the Foundations of Data Science Institute.

    He has previously received such honors as the 2018 Nevanlinna Prize from the International Mathematical Union, the 2018 ACM Grace Murray Hopper Award, the Kalai Game Theory and Computer Science Prize from the Game Theory Society, and the 2008 ACM Doctoral Dissertation Award. More

  • in

    Teaching AI to ask clinical questions

    Physicians often query a patient’s electronic health record for information that helps them make treatment decisions, but the cumbersome nature of these records hampers the process. Research has shown that even when a doctor has been trained to use an electronic health record (EHR), finding an answer to just one question can take, on average, more than eight minutes.

    The more time physicians must spend navigating an oftentimes clunky EHR interface, the less time they have to interact with patients and provide treatment.

    Researchers have begun developing machine-learning models that can streamline the process by automatically finding information physicians need in an EHR. However, training effective models requires huge datasets of relevant medical questions, which are often hard to come by due to privacy restrictions. Existing models struggle to generate authentic questions — those that would be asked by a human doctor — and are often unable to successfully find correct answers.

    To overcome this data shortage, researchers at MIT partnered with medical experts to study the questions physicians ask when reviewing EHRs. Then, they built a publicly available dataset of more than 2,000 clinically relevant questions written by these medical experts.

    When they used their dataset to train a machine-learning model to generate clinical questions, they found that the model asked high-quality and authentic questions, as compared to real questions from medical experts, more than 60 percent of the time.

    With this dataset, they plan to generate vast numbers of authentic medical questions and then use those questions to train a machine-learning model which would help doctors find sought-after information in a patient’s record more efficiently.

    “Two thousand questions may sound like a lot, but when you look at machine-learning models being trained nowadays, they have so much data, maybe billions of data points. When you train machine-learning models to work in health care settings, you have to be really creative because there is such a lack of data,” says lead author Eric Lehman, a graduate student in the Computer Science and Artificial Intelligence Laboratory (CSAIL).

    The senior author is Peter Szolovits, a professor in the Department of Electrical Engineering and Computer Science (EECS) who heads the Clinical Decision-Making Group in CSAIL and is also a member of the MIT-IBM Watson AI Lab. The research paper, a collaboration between co-authors at MIT, the MIT-IBM Watson AI Lab, IBM Research, and the doctors and medical experts who helped create questions and participated in the study, will be presented at the annual conference of the North American Chapter of the Association for Computational Linguistics.

    “Realistic data is critical for training models that are relevant to the task yet difficult to find or create,” Szolovits says. “The value of this work is in carefully collecting questions asked by clinicians about patient cases, from which we are able to develop methods that use these data and general language models to ask further plausible questions.”

    Data deficiency

    The few large datasets of clinical questions the researchers were able to find had a host of issues, Lehman explains. Some were composed of medical questions asked by patients on web forums, which are a far cry from physician questions. Other datasets contained questions produced from templates, so they are mostly identical in structure, making many questions unrealistic.

    “Collecting high-quality data is really important for doing machine-learning tasks, especially in a health care context, and we’ve shown that it can be done,” Lehman says.

    To build their dataset, the MIT researchers worked with practicing physicians and medical students in their last year of training. They gave these medical experts more than 100 EHR discharge summaries and told them to read through a summary and ask any questions they might have. The researchers didn’t put any restrictions on question types or structures in an effort to gather natural questions. They also asked the medical experts to identify the “trigger text” in the EHR that led them to ask each question.

    For instance, a medical expert might read a note in the EHR that says a patient’s past medical history is significant for prostate cancer and hypothyroidism. The trigger text “prostate cancer” could lead the expert to ask questions like “date of diagnosis?” or “any interventions done?”

    They found that most questions focused on symptoms, treatments, or the patient’s test results. While these findings weren’t unexpected, quantifying the number of questions about each broad topic will help them build an effective dataset for use in a real, clinical setting, says Lehman.

    Once they had compiled their dataset of questions and accompanying trigger text, they used it to train machine-learning models to ask new questions based on the trigger text.

    Then the medical experts determined whether those questions were “good” using four metrics: understandability (Does the question make sense to a human physician?), triviality (Is the question too easily answerable from the trigger text?), medical relevance (Does it makes sense to ask this question based on the context?), and relevancy to the trigger (Is the trigger related to the question?).

    Cause for concern

    The researchers found that when a model was given trigger text, it was able to generate a good question 63 percent of the time, whereas a human physician would ask a good question 80 percent of the time.

    They also trained models to recover answers to clinical questions using the publicly available datasets they had found at the outset of this project. Then they tested these trained models to see if they could find answers to “good” questions asked by human medical experts.

    The models were only able to recover about 25 percent of answers to physician-generated questions.

    “That result is really concerning. What people thought were good-performing models were, in practice, just awful because the evaluation questions they were testing on were not good to begin with,” Lehman says.

    The team is now applying this work toward their initial goal: building a model that can automatically answer physicians’ questions in an EHR. For the next step, they will use their dataset to train a machine-learning model that can automatically generate thousands or millions of good clinical questions, which can then be used to train a new model for automatic question answering.

    While there is still much work to do before that model could be a reality, Lehman is encouraged by the strong initial results the team demonstrated with this dataset.

    This research was supported, in part, by the MIT-IBM Watson AI Lab. Additional co-authors include Leo Anthony Celi of the MIT Institute for Medical Engineering and Science; Preethi Raghavan and Jennifer J. Liang of the MIT-IBM Watson AI Lab; Dana Moukheiber of the University of Buffalo; Vladislav Lialin and Anna Rumshisky of the University of Massachusetts at Lowell; Katelyn Legaspi, Nicole Rose I. Alberto, Richard Raymund R. Ragasa, Corinna Victoria M. Puyat, Isabelle Rose I. Alberto, and Pia Gabrielle I. Alfonso of the University of the Philippines; Anne Janelle R. Sy and Patricia Therese S. Pile of the University of the East Ramon Magsaysay Memorial Medical Center; Marianne Taliño of the Ateneo de Manila University School of Medicine and Public Health; and Byron C. Wallace of Northeastern University. More

  • in

    Robots play with play dough

    The inner child in many of us feels an overwhelming sense of joy when stumbling across a pile of the fluorescent, rubbery mixture of water, salt, and flour that put goo on the map: play dough. (Even if this happens rarely in adulthood.)

    While manipulating play dough is fun and easy for 2-year-olds, the shapeless sludge is hard for robots to handle. Machines have become increasingly reliable with rigid objects, but manipulating soft, deformable objects comes with a laundry list of technical challenges, and most importantly, as with most flexible structures, if you move one part, you’re likely affecting everything else. 

    Scientists from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and Stanford University recently let robots take their hand at playing with the modeling compound, but not for nostalgia’s sake. Their new system learns directly from visual inputs to let a robot with a two-fingered gripper see, simulate, and shape doughy objects. “RoboCraft” could reliably plan a robot’s behavior to pinch and release play dough to make various letters, including ones it had never seen. With just 10 minutes of data, the two-finger gripper rivaled human counterparts that teleoperated the machine — performing on-par, and at times even better, on the tested tasks. 

    “Modeling and manipulating objects with high degrees of freedom are essential capabilities for robots to learn how to enable complex industrial and household interaction tasks, like stuffing dumplings, rolling sushi, and making pottery,” says Yunzhu Li, CSAIL PhD student and author on a new paper about RoboCraft. “While there’s been recent advances in manipulating clothes and ropes, we found that objects with high plasticity, like dough or plasticine — despite ubiquity in those household and industrial settings — was a largely underexplored territory. With RoboCraft, we learn the dynamics models directly from high-dimensional sensory data, which offers a promising data-driven avenue for us to perform effective planning.” 

    Play video

    With undefined, smooth material, the whole structure needs to be accounted for before you can do any type of efficient and effective modeling and planning. By turning the images into graphs of little particles, coupled with algorithms, RoboCraft, using a graph neural network as the dynamics model, makes more accurate predictions about the material’s change of shapes. 

    Typically, researchers have used complex physics simulators to model and understand force and dynamics being applied to objects, but RoboCraft simply uses visual data. The inner-workings of the system relies on three parts to shape soft material into, say, an “R.” 

    The first part — perception — is all about learning to “see.” It uses cameras to collect raw, visual sensor data from the environment, which are then turned into little clouds of particles to represent the shapes. A graph-based neural network then uses said particle data to learn to “simulate” the object’s dynamics, or how it moves. Then, algorithms help plan the robot’s behavior so it learns to “shape” a blob of dough, armed with the training data from the many pinches. While the letters are a bit loose, they’re indubitably representative. 

    Besides cutesy shapes, the team is (actually) working on making dumplings from dough and a prepared filling. Right now, with just a two finger gripper, it’s a big ask. RoboCraft would need additional tools (a baker needs multiple tools to cook; so do robots) — a rolling pin, a stamp, and a mold. 

    A more far in the future domain the scientists envision is using RoboCraft for assistance with household tasks and chores, which could be of particular help to the elderly or those with limited mobility. To accomplish this, given the many obstructions that could take place, a much more adaptive representation of the dough or item would be needed, and as well as exploration into what class of models might be suitable to capture the underlying structural systems. 

    “RoboCraft essentially demonstrates that this predictive model can be learned in very data-efficient ways to plan motion. In the long run, we are thinking about using various tools to manipulate materials,” says Li. “If you think about dumpling or dough making, just one gripper wouldn’t be able to solve it. Helping the model understand and accomplish longer-horizon planning tasks, such as, how the dough will deform given the current tool, movements and actions, is a next step for future work.” 

    Li wrote the paper alongside Haochen Shi, Stanford master’s student; Huazhe Xu, Stanford postdoc; Zhiao Huang, PhD student at the University of California at San Diego; and Jiajun Wu, assistant professor at Stanford. They will present the research at the Robotics: Science and Systems conference in New York City. The work is in part supported by the Stanford Institute for Human-Centered AI (HAI), the Samsung Global Research Outreach (GRO) Program, the Toyota Research Institute (TRI), and Amazon, Autodesk, Salesforce, and Bosch. More

  • in

    Researchers release open-source photorealistic simulator for autonomous driving

    Hyper-realistic virtual worlds have been heralded as the best driving schools for autonomous vehicles (AVs), since they’ve proven fruitful test beds for safely trying out dangerous driving scenarios. Tesla, Waymo, and other self-driving companies all rely heavily on data to enable expensive and proprietary photorealistic simulators, since testing and gathering nuanced I-almost-crashed data usually isn’t the most easy or desirable to recreate. 

    To that end, scientists from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) created “VISTA 2.0,” a data-driven simulation engine where vehicles can learn to drive in the real world and recover from near-crash scenarios. What’s more, all of the code is being open-sourced to the public. 

    “Today, only companies have software like the type of simulation environments and capabilities of VISTA 2.0, and this software is proprietary. With this release, the research community will have access to a powerful new tool for accelerating the research and development of adaptive robust control for autonomous driving,” says MIT Professor and CSAIL Director Daniela Rus, senior author on a paper about the research. 

    Play video

    VISTA is a data-driven, photorealistic simulator for autonomous driving. It can simulate not just live video but LiDAR data and event cameras, and also incorporate other simulated vehicles to model complex driving situations. VISTA is open source and the code can be found below.

    VISTA 2.0 builds off of the team’s previous model, VISTA, and it’s fundamentally different from existing AV simulators since it’s data-driven — meaning it was built and photorealistically rendered from real-world data — thereby enabling direct transfer to reality. While the initial iteration supported only single car lane-following with one camera sensor, achieving high-fidelity data-driven simulation required rethinking the foundations of how different sensors and behavioral interactions can be synthesized. 

    Enter VISTA 2.0: a data-driven system that can simulate complex sensor types and massively interactive scenarios and intersections at scale. With much less data than previous models, the team was able to train autonomous vehicles that could be substantially more robust than those trained on large amounts of real-world data. 

    “This is a massive jump in capabilities of data-driven simulation for autonomous vehicles, as well as the increase of scale and ability to handle greater driving complexity,” says Alexander Amini, CSAIL PhD student and co-lead author on two new papers, together with fellow PhD student Tsun-Hsuan Wang. “VISTA 2.0 demonstrates the ability to simulate sensor data far beyond 2D RGB cameras, but also extremely high dimensional 3D lidars with millions of points, irregularly timed event-based cameras, and even interactive and dynamic scenarios with other vehicles as well.” 

    The team was able to scale the complexity of the interactive driving tasks for things like overtaking, following, and negotiating, including multiagent scenarios in highly photorealistic environments. 

    Training AI models for autonomous vehicles involves hard-to-secure fodder of different varieties of edge cases and strange, dangerous scenarios, because most of our data (thankfully) is just run-of-the-mill, day-to-day driving. Logically, we can’t just crash into other cars just to teach a neural network how to not crash into other cars.

    Recently, there’s been a shift away from more classic, human-designed simulation environments to those built up from real-world data. The latter have immense photorealism, but the former can easily model virtual cameras and lidars. With this paradigm shift, a key question has emerged: Can the richness and complexity of all of the sensors that autonomous vehicles need, such as lidar and event-based cameras that are more sparse, accurately be synthesized? 

    Lidar sensor data is much harder to interpret in a data-driven world — you’re effectively trying to generate brand-new 3D point clouds with millions of points, only from sparse views of the world. To synthesize 3D lidar point clouds, the team used the data that the car collected, projected it into a 3D space coming from the lidar data, and then let a new virtual vehicle drive around locally from where that original vehicle was. Finally, they projected all of that sensory information back into the frame of view of this new virtual vehicle, with the help of neural networks. 

    Together with the simulation of event-based cameras, which operate at speeds greater than thousands of events per second, the simulator was capable of not only simulating this multimodal information, but also doing so all in real time — making it possible to train neural nets offline, but also test online on the car in augmented reality setups for safe evaluations. “The question of if multisensor simulation at this scale of complexity and photorealism was possible in the realm of data-driven simulation was very much an open question,” says Amini. 

    With that, the driving school becomes a party. In the simulation, you can move around, have different types of controllers, simulate different types of events, create interactive scenarios, and just drop in brand new vehicles that weren’t even in the original data. They tested for lane following, lane turning, car following, and more dicey scenarios like static and dynamic overtaking (seeing obstacles and moving around so you don’t collide). With the multi-agency, both real and simulated agents interact, and new agents can be dropped into the scene and controlled any which way. 

    Taking their full-scale car out into the “wild” — a.k.a. Devens, Massachusetts — the team saw  immediate transferability of results, with both failures and successes. They were also able to demonstrate the bodacious, magic word of self-driving car models: “robust.” They showed that AVs, trained entirely in VISTA 2.0, were so robust in the real world that they could handle that elusive tail of challenging failures. 

    Now, one guardrail humans rely on that can’t yet be simulated is human emotion. It’s the friendly wave, nod, or blinker switch of acknowledgement, which are the type of nuances the team wants to implement in future work. 

    “The central algorithm of this research is how we can take a dataset and build a completely synthetic world for learning and autonomy,” says Amini. “It’s a platform that I believe one day could extend in many different axes across robotics. Not just autonomous driving, but many areas that rely on vision and complex behaviors. We’re excited to release VISTA 2.0 to help enable the community to collect their own datasets and convert them into virtual worlds where they can directly simulate their own virtual autonomous vehicles, drive around these virtual terrains, train autonomous vehicles in these worlds, and then can directly transfer them to full-sized, real self-driving cars.” 

    Amini and Wang wrote the paper alongside Zhijian Liu, MIT CSAIL PhD student; Igor Gilitschenski, assistant professor in computer science at the University of Toronto; Wilko Schwarting, AI research scientist and MIT CSAIL PhD ’20; Song Han, associate professor at MIT’s Department of Electrical Engineering and Computer Science; Sertac Karaman, associate professor of aeronautics and astronautics at MIT; and Daniela Rus, MIT professor and CSAIL director. The researchers presented the work at the IEEE International Conference on Robotics and Automation (ICRA) in Philadelphia. 

    This work was supported by the National Science Foundation and Toyota Research Institute. The team acknowledges the support of NVIDIA with the donation of the Drive AGX Pegasus. More

  • in

    Hallucinating to better text translation

    As babies, we babble and imitate our way to learning languages. We don’t start off reading raw text, which requires fundamental knowledge and understanding about the world, as well as the advanced ability to interpret and infer descriptions and relationships. Rather, humans begin our language journey slowly, by pointing and interacting with our environment, basing our words and perceiving their meaning through the context of the physical and social world. Eventually, we can craft full sentences to communicate complex ideas.

    Similarly, when humans begin learning and translating into another language, the incorporation of other sensory information, like multimedia, paired with the new and unfamiliar words, like flashcards with images, improves language acquisition and retention. Then, with enough practice, humans can accurately translate new, unseen sentences in context without the accompanying media; however, imagining a picture based on the original text helps.

    This is the basis of a new machine learning model, called VALHALLA, by researchers from MIT, IBM, and the University of California at San Diego, in which a trained neural network sees a source sentence in one language, hallucinates an image of what it looks like, and then uses both to translate into a target language. The team found that their method demonstrates improved accuracy of machine translation over text-only translation. Further, it provided an additional boost for cases with long sentences, under-resourced languages, and instances where part of the source sentence is inaccessible to the machine translator.

    As a core task within the AI field of natural language processing (NLP), machine translation is an “eminently practical technology that’s being used by millions of people every day,” says study co-author Yoon Kim, assistant professor in MIT’s Department of Electrical Engineering and Computer Science with affiliations in the Computer Science and Artificial Intelligence Laboratory (CSAIL) and the MIT-IBM Watson AI Lab. With recent, significant advances in deep learning, “there’s been an interesting development in how one might use non-text information — for example, images, audio, or other grounding information — to tackle practical tasks involving language” says Kim, because “when humans are performing language processing tasks, we’re doing so within a grounded, situated world.” The pairing of hallucinated images and text during inference, the team postulated, imitates that process, providing context for improved performance over current state-of-the-art techniques, which utilize text-only data.

    This research will be presented at the IEEE / CVF Computer Vision and Pattern Recognition Conference this month. Kim’s co-authors are UC San Diego graduate student Yi Li and Professor Nuno Vasconcelos, along with research staff members Rameswar Panda, Chun-fu “Richard” Chen, Rogerio Feris, and IBM Director David Cox of IBM Research and the MIT-IBM Watson AI Lab.

    Learning to hallucinate from images

    When we learn new languages and to translate, we’re often provided with examples and practice before venturing out on our own. The same is true for machine-translation systems; however, if images are used during training, these AI methods also require visual aids for testing, limiting their applicability, says Panda.

    “In real-world scenarios, you might not have an image with respect to the source sentence. So, our motivation was basically: Instead of using an external image during inference as input, can we use visual hallucination — the ability to imagine visual scenes — to improve machine translation systems?” says Panda.

    To do this, the team used an encoder-decoder architecture with two transformers, a type of neural network model that’s suited for sequence-dependent data, like language, that can pay attention key words and semantics of a sentence. One transformer generates a visual hallucination, and the other performs multimodal translation using outputs from the first transformer.

    During training, there are two streams of translation: a source sentence and a ground-truth image that is paired with it, and the same source sentence that is visually hallucinated to make a text-image pair. First the ground-truth image and sentence are tokenized into representations that can be handled by transformers; for the case of the sentence, each word is a token. The source sentence is tokenized again, but this time passed through the visual hallucination transformer, outputting a hallucination, a discrete image representation of the sentence. The researchers incorporated an autoregression that compares the ground-truth and hallucinated representations for congruency — e.g., homonyms: a reference to an animal “bat” isn’t hallucinated as a baseball bat. The hallucination transformer then uses the difference between them to optimize its predictions and visual output, making sure the context is consistent.

    The two sets of tokens are then simultaneously passed through the multimodal translation transformer, each containing the sentence representation and either the hallucinated or ground-truth image. The tokenized text translation outputs are compared with the goal of being similar to each other and to the target sentence in another language. Any differences are then relayed back to the translation transformer for further optimization.

    For testing, the ground-truth image stream drops off, since images likely wouldn’t be available in everyday scenarios.

    “To the best of our knowledge, we haven’t seen any work which actually uses a hallucination transformer jointly with a multimodal translation system to improve machine translation performance,” says Panda.

    Visualizing the target text

    To test their method, the team put VALHALLA up against other state-of-the-art multimodal and text-only translation methods. They used public benchmark datasets containing ground-truth images with source sentences, and a dataset for translating text-only news articles. The researchers measured its performance over 13 tasks, ranging from translation on well-resourced languages (like English, German, and French), under-resourced languages (like English to Romanian) and non-English (like Spanish to French). The group also tested varying transformer model sizes, how accuracy changes with the sentence length, and translation under limited textual context, where portions of the text were hidden from the machine translators.

    The team observed significant improvements over text-only translation methods, improving data efficiency, and that smaller models performed better than the larger base model. As sentences became longer, VALHALLA’s performance over other methods grew, which the researchers attributed to the addition of more ambiguous words. In cases where part of the sentence was masked, VALHALLA could recover and translate the original text, which the team found surprising.

    Further unexpected findings arose: “Where there weren’t as many training [image and] text pairs, [like for under-resourced languages], improvements were more significant, which indicates that grounding in images helps in low-data regimes,” says Kim. “Another thing that was quite surprising to me was this improved performance, even on types of text that aren’t necessarily easily connectable to images. For example, maybe it’s not so surprising if this helps in translating visually salient sentences, like the ‘there is a red car in front of the house.’ [However], even in text-only [news article] domains, the approach was able to improve upon text-only systems.”

    While VALHALLA performs well, the researchers note that it does have limitations, requiring pairs of sentences to be annotated with an image, which could make it more expensive to obtain. It also performs better in its ground domain and not the text-only news articles. Moreover, Kim and Panda note, a technique like VALHALLA is still a black box, with the assumption that hallucinated images are providing helpful information, and the team plans to investigate what and how the model is learning in order to validate their methods.

    In the future, the team plans to explore other means of improving translation. “Here, we only focus on images, but there are other types of a multimodal information — for example, speech, video or touch, or other sensory modalities,” says Panda. “We believe such multimodal grounding can lead to even more efficient machine translation models, potentially benefiting translation across many low-resource languages spoken in the world.”

    This research was supported, in part, by the MIT-IBM Watson AI Lab and the National Science Foundation. More